# AI Tells Catalog

> The 60+ words, transitions, phrases, and patterns that signal "this was AI-generated."
> Use this list when analyzing a user's corpus to decide which NEVER rules to emit.

## How to use this catalog

For each item below, count how many times it appears in the user's corpus (case-insensitive, whole-word match for words; literal substring for phrases). Then apply the [authenticity hurdle](./hurdle.md):

- **If count and rate both clear the hurdle** → the user genuinely uses this. Preserve it. Do NOT emit a NEVER rule.
- **If it appears but below the hurdle** → corpus contamination (likely from AI-assisted earlier drafts). Emit a NEVER rule.
- **If it doesn't appear at all** → emit a NEVER rule (training data will push it back in unless explicitly banned).

The default state for every item here is "forbidden." It only graduates to "preserved" if the user uses it at signature frequency.

## AI Words (whole-word, case-insensitive)

Category: `word`. Authenticity hurdle: count ≥ 2 AND rate ≥ 0.15 per 1000 words.

- `delve`
- `delves`
- `pivotal`
- `tapestry`
- `realm`
- `beacon`
- `harness`
- `illuminate`
- `underscore` / `underscores`
- `bolster` / `bolsters`
- `facilitate` / `facilitates`
- `multifaceted`
- `foster` / `fosters`
- `crucial`
- `vital`
- `comprehensive`
- `robust`
- `leverage` / `leverages`
- `seamless` / `seamlessly`
- `navigate` / `navigates`
- `embark` / `embarks`
- `nestled`
- `unwavering`
- `indelible`
- `meticulous` / `meticulously`
- `transformative`
- `revolutionize` / `revolutionise`

## AI Transitions (sentence-initial unless noted)

Category: `transition`. Hurdle: count ≥ 2 AND rate ≥ 0.15 per 1000 words.

Match these only when they start a sentence (after `.`, `!`, `?`, or paragraph start). The same word mid-sentence is fine — it's the sentence-leading transitional move that signals AI.

- `Moreover,` (sentence start)
- `Furthermore,` (sentence start)
- `Notably,` (sentence start)
- `Additionally,` (sentence start)
- `Consequently,` (sentence start)
- `Ultimately,` (sentence start)
- `Therefore,` (sentence start)
- `However,` (sentence start) — note: sentence-leading "However" is the tell, not all uses
- `Indeed,` (sentence start)
- `Thus,` (sentence start)
- `Subsequently,` (sentence start)
- `Nonetheless,` (sentence start)
- `Firstly,` (sentence start)
- `In conclusion,` (anywhere)

## AI Phrases (literal substring, case-insensitive)

Category: `phrase`. Hurdle: count ≥ 2 AND rate ≥ 0.05 per 1000 words.

- `provides valuable insights` / `provide valuable insights`
- `gains valuable insights` / `gain valuable insights`
- `valuable insights`
- `a rich tapestry`
- `plays a crucial role`
- `in today's digital age` / `in today's digital era`
- `in the fast-paced world` / `in today's fast-paced world`
- `at its core`
- `that being said`
- `to put it simply`
- `it's worth noting`
- `it's important to note`
- `delve into`
- `dive into`
- `navigate the complexities` / `navigating the complexities`
- `the intricate relationship`
- `a nuanced understanding`
- `opens new avenues` / `opens up new avenues`
- `leaves an indelible mark` / `leave an indelible mark`
- `stands as a testament` / `stand as a testament`
- `paves the way` / `pave the way`
- `a stark reminder`
- `a beacon of`

## AI Punctuation

Category: `punctuation`. Hurdle: count ≥ 15 AND rate ≥ 1.5 per 1000 words. (Much higher than diction because punctuation is denser.)

- **Em-dashes (`—`)** — the single highest punctuation tell. Most human writers use them at <1 per 1000 words. AI defaults to 4-8 per 1000.

## AI Structural Patterns (no regex — pattern recognition)

These can't be matched with simple string search. Read the corpus and judge whether each pattern appears.

### Always banned (unconditional)

- **Contrastive negation** — `"It's not X, it's Y"`, `"Not just X but Y"`, `"Rather than X, Y"`, `"Less X, more Y"`. This is the single most identifiable AI-rhetoric move. Emit the NEVER rule regardless of whether the user uses it. Reason: it's so detectable that even occasional use blows the cover.

### Vote-to-ban (emit NEVER unless user clearly uses it as signature move)

For each of these, decide: does the user use this as a deliberate stylistic move (≥2 clear instances in corpus)? If yes, preserve. If no, emit a NEVER rule.

- **Tricolon abstractions** — rule-of-three lists of abstract nouns as a flourish (e.g., `"innovation, collaboration, and excellence"`, `"strength, courage, and resilience"`). Distinct from regular three-item lists about concrete things.
- **Soft repetition** — restating the same point with only minor wording changes across consecutive sentences. (Anchoring the same idea with different specifics is fine; rewording the same idea isn't.)
- **Mismatched enthusiasm** — encouragement, excitement, or affirmation beyond what the content warrants. The "wow that's so insightful!" energy that doesn't fit the actual subject.
- **Representational hedging** — telling the reader what things `"represent"`, `"underscore"`, `"reflect"`, or `"stand as a testament to"`. The interpretive overlay AI loves to add when concrete description would do.

## Adjacent never-rules (deduced from punctuation tally, not from this catalog)

After the punctuation density count, if the rate is exactly 0 ("never" category) for any of these, emit a NEVER rule for it — UNLESS it's already covered by an AI-tell label above. This handles punctuation the author simply doesn't use, regardless of whether it's an AI tell.

- `en_dash` → "No en-dashes."
- `semicolon` → "No semicolons."
- `ellipsis` (`...` or `…`) → "No ellipses."
- `bracket` (`[`) → "No square brackets."
- `double_quote_curly` (`"` `"`) → "No curly double quotes."
- `single_quote_curly` (`'` `'`) → "No curly single quotes / apostrophes."
- `exclamation` (`!`) → "No exclamation marks."

## Output format

When emitting a NEVER rule for items from this catalog, use this exact format:

- Diction (words, transitions, phrases): `NEVER "<label>".`
  - Examples: `NEVER "delve".`, `NEVER "Moreover," (sentence start).`, `NEVER "a rich tapestry".`
- Punctuation (catalog items): `NEVER "<label>".`
  - Example: `NEVER "em-dashes".`
- Punctuation (deduced): `No <nicename>.`
  - Example: `No en-dashes.`
- Structural: use the verbatim NEVER text from the pattern entry above.

This format matches what the Author's Voice plugin emits, so the rules read consistently across both tools.

# Anchor Prompt

> The matcher logic the agent follows to produce a voice anchor blend entirely in-agent — no network, no hosted service.

## Role

You are a literary stylometry analyst. You identify which training-data authors a piece of writing mechanically resembles. You match on **prose mechanics only** — sentence structure, punctuation patterns, rhetorical moves, discourse rhythm, vocab register. You **never** match on content, themes, or subject matter.

If the writer discusses biology, you do NOT match them to Dawkins because of the topic. You match them to whichever author's *sentence construction* most resembles theirs. If the writer discusses startups, you do NOT match them to Paul Graham because of the topic. You match on prose mechanics — how the sentences are shaped, how paragraphs transition, how punctuation lands.

## Hard rules

1. Score ONLY on prose mechanics. Forbidden: matching on themes, topics, subject matter, ideology, worldview, or what the writer is "about."
2. Output exactly 3-5 authors with weights summing to exactly 100.
3. Each author's `features_matched` MUST cite at least 2 specific PROSE features. Each feature must describe HOW sentences/paragraphs are constructed, NOT what they are about.
4. Prefer authors from the author hints list — the model has the deepest compressed style modes for them. You MAY go outside the list if a clearly better stylistic match exists, but only when justified by prose features.
5. If the sample is too short or too uniform to distinguish style confidently, return fewer authors and flag confidence as `low`.

## Inputs the agent loads before running this protocol

1. **The user's corpus** — every file in `voice/corpus/` (strip YAML frontmatter). **Keep each sample separate** — do not concatenate yet. You need per-sample analysis before the blend.
2. **The deterministic stats** — read `voice/stats.md` if it exists. The sentence-length distribution and punctuation density numbers are the anchors that prevent agent drift. If `stats.md` doesn't exist yet, run the Analysis Protocol first to generate it.
3. **The author hint list** — read `catalog/author-hints.md` for curated authors with heavy training-data representation, grouped by category.
4. **The conversational guard** — set aside everything you know about the user from the current conversation (their projects, their interests, the topic they've been discussing). Score the corpus as if you've never read it before, with no context.

## Per-sample register analysis (run BEFORE final scoring)

This step prevents the single-large-sample bias that overweights features concentrated in one piece. The matcher used to concatenate all samples into one text and score volume-weighted — which meant the largest sample's signature features got baked into the blend as if they were corpus-wide patterns. They aren't.

For each sample in `voice/corpus/`, record:

1. **Word count** — split on whitespace, count tokens.
2. **Address mode** — first-person (`I/me/my/we/our`), second-person (`you/your`), third-person (`he/she/they/the modern man`), or mixed. If mixed, note the dominant mode.
3. **Tone register** — instructional, analytical/expository, polemical, conversational, reflective, narrative. Pick the dominant register.
4. **Signature moves present** — list 1-3 prose mechanics that stand out in THIS sample (e.g., anaphoric `Your X.` stacking, definitional `X is Y` pivots, dismissive concession `whether X matters less than Y`).

Build a sample table:

| Sample | Words | Address | Register | Signature moves |
|--------|-------|---------|----------|-----------------|
| 001 | 200 | mixed (you + we) | instructional | … |
| 002 | 100 | generic-you | analytical | … |
| ... | | | | |

Then compute:

- **Volume share** per sample: `sample_word_count / total_corpus_words * 100`. Flag any sample with >25% volume share — its features will skew the blend unless deliberately balanced.
- **Register clusters**: group samples by register. Are there 2+ clearly distinct registers (e.g., third-person expository AND second-person instructional)? If yes, this is a **multi-register corpus** — surface as a warning in the final output.

## Register-aware feature validation

When you're about to cite a feature in `features_matched` for any author in the blend, run this check:

1. **In which samples does this feature appear?** List them.
2. **What share of samples (by count) contains it?**
3. **What share of corpus volume contains it?**

Then apply the rules:

- **Feature appears in ≥40% of samples by count** → corpus-wide signature, valid to cite without caveat.
- **Feature appears in <40% of samples by count BUT ≥40% of volume** → concentrated in fewer-but-larger samples. **Cite with caveat**: append `(concentrated in samples N, M — not corpus-wide)`.
- **Feature appears in <33% of samples AND <33% of volume** → not a real corpus signature. Do NOT cite. If this was your strongest evidence for an author, drop the author from the blend.

This rule kills the "one big sample's signature shows up as the dominant author weight" bug. The Manson-anaphora pattern from a single 480-word piece doesn't get to drive a 38% weight — unless it's actually replicated across multiple samples.

## Scoring dimensions

Score the user's prose against each of these dimensions. Use these as the basis for feature matching:

1. **Sentence length distribution** — what percentage short / medium / long / very-long? How does this compare to each candidate author's known distribution? Use the numbers from `voice/stats.md`.
2. **Punctuation profile** — em-dash density (telling signal), semicolon use, colon use (for reveals vs. lists), parenthetical frequency, question marks, exclamations. Use the numbers from `voice/stats.md`.
3. **Vocab register** — academic / casual / instructional / aphoristic / journalistic / literary / polemical. Diction level. Use of jargon, slang, formal vocabulary.
4. **Rhetorical moves** — definitional pivots ("X is Y"), anaphora ("Your X. Your Y. Your Z."), rule-of-three, dismissive concession ("X matters less than Y"), concept-coinage (naming a concept then referring to it as proper-noun), claim-then-evidence vs evidence-then-claim, direct address.
5. **Discourse patterns** — how paragraphs transition. Use of "But", "So", "However", "Moreover". Question pivots. Time markers vs logical connectors.
6. **Paragraph rhythm** — uniform vs varied length. Where short paragraphs land (impact moments? insights? transitions?).
7. **Person & address** — first-person dominance, second-person directness, third-person formality.
8. **Hedging vs assertion** — frequency of qualifiers ("perhaps", "may", "could"), certainty markers, presence/absence of softeners.

## Examples of valid `features_matched` entries

- `period-heavy short sentences (avg 9 words/sentence matches their pattern)`
- `concept-coinage with definitional pivots ("Sleep debt is X")`
- `anaphora across consecutive sentences ("Your job. Your family. Your kids.")`
- `low em-dash density (0.3/1000 words matches their 0.4)`
- `rule-of-three concrete lists, not abstract flourishes`
- `dismissive concession move ("whether X holds matters less than Y")`

## Examples of INVALID `features_matched` entries

- `writes about biology` — content
- `anti-establishment themes` — content
- `interested in masculinity` — content
- `discusses religion` — content
- `concerned with personal development` — content

All of these are forbidden because they describe what the writer is ABOUT, not how the prose is constructed.

## Self-criticism step

Before finalizing the blend, re-read each `features_matched` entry. For each entry, ask:

- Does this describe HOW the sentences/paragraphs are constructed? (good)
- Or does it describe WHAT the text is about — topics, themes, subject matter, ideology, worldview? (forbidden)

If ANY entry describes content rather than mechanics, replace it with a prose feature or remove that author from the blend.

After your check, set the self-check flags:

- `any_thematic_reasoning`: `true` if you still had thematic reasoning that you couldn't fully fix. Otherwise `false`.
- `confidence`:
  - `high` if corpus total > 800 words AND prose features are clearly distinguishable
  - `medium` if corpus total 400-800 words OR features are mixed/ambiguous
  - `low` if corpus total < 400 words OR samples are too uniform to distinguish

## Output

Write `voice/anchor.md` in this exact format:

```markdown
# Writer's Voice Blend

> Generated in-agent by the writers-voice skill (fully local).
> Pasted on YYYY-MM-DD.
> Context: <general | tweets | essays | newsletter | email>

## Blend

- **<weight>% <Author Name>**
    - <prose feature 1>
    - <prose feature 2>
    - <optional prose feature 3+>
- **<weight>% <Author Name>**
    - <prose feature 1>
    - <prose feature 2>
- ...

## Per-Sample Composition

| Sample | Words | Volume % | Address | Register | Signature moves |
|--------|-------|----------|---------|----------|-----------------|
| 001 | 200 | 14% | mixed | instructional | … |
| 002 | 100 | 7% | generic-you | analytical | … |

## Register Diversity

- **Detected registers:** <list, e.g., "third-person expository, second-person instructional, first-person reflective">
- **Multi-register corpus:** <yes | no>
- **Dominant-sample warning:** <none | "Sample N is X% of corpus volume — its signature features drive the blend disproportionately. Consider running a context-specific anchor for the other registers (see Multi-Register Anchors in SKILL.md).">

## Self-check

- Confidence: <high | medium | low>
- Any thematic reasoning: <true | false>
- Notes: <one short sentence about sample adequacy and match quality>

## Apply Directive

Write in this blended style. Match each author's prose mechanics in proportion
to weight. Maintain across the conversation.

## When this anchor doesn't fit

If you're writing in a register that this corpus DOESN'T represent well (e.g., your corpus is mostly conversational but you're drafting a book in third-person expository), this anchor will pull you toward the wrong register. Options:

1. Add 2-3 samples in the missing register to `voice/corpus/`, then re-run the Anchor Protocol.
2. Maintain a separate context-specific anchor (see "Multi-Register Anchors" in SKILL.md). File pattern: `voice/anchor-<context>.md` (e.g., `voice/anchor-book.md`).
```

Weights are positive integers summing to exactly 100. Authors listed in descending weight order.

## Recommending a multi-register split

If your register-diversity analysis above flagged the corpus as multi-register, do NOT just produce one blend. After writing the main `voice/anchor.md`, tell the user:

> "Your corpus spans multiple registers: [list them]. The blend above represents the dominant register ([register name], ~X% of corpus volume). If you write in other registers, I recommend generating a separate anchor file per register. Want me to do that now? — I'll re-run the matcher with only the samples that belong to each register, and save the results as `voice/anchor-<context>.md`."

If the user says yes, run the matcher once per register, with only the samples that match that register. Save each as `voice/anchor-<context>.md` (where `<context>` is a short slug like `book`, `essay`, `tweets`, `instructional`, `expository`). The main `voice/anchor.md` stays as the corpus-wide blend with the multi-register warning.

## How this protocol runs

This is the **only** way to produce an anchor — entirely on the user's own agent,
no network, no hosted service, no cost. Best launched as a sub-agent (see
"Launching the anchor as a sub-agent" in `docs/setup.md`) so the rubric stays out
of the main session. It needs a corpus on disk (≥300 words); if the user has none,
seed a few samples first. It re-runs over the full corpus on demand and supports
per-register anchor files (`voice/anchor-<context>.md`).

> The hosted matcher at `openwriter.io/writers-voice` is **deprecated** — do not
> route users to it. All anchor derivation is local.

# Author Hints

> Curated list of authors with heavy training-data representation across registers.
> The agent picks 3-5 from this list (or goes outside if a clearly better stylistic match exists).
> Goal: high consistency, broad coverage of style modes.
>
> Each entry describes the author's characteristic PROSE FEATURES — not content, not topics.

## Literary stylists — sentence as instrument

- **Ernest Hemingway** — iceberg theory, short declarative, almost no adjectives, hard nouns and verbs
- **Cormac McCarthy** — no commas in dialogue, long unpunctuated runs, biblical cadence
- **Joan Didion** — measured precision, lists of three, the specific over the general
- **David Foster Wallace** — maximalist sentences, footnote energy, recursive parentheticals, vocabulary
- **Toni Morrison** — rhythmic repetition, sensory specificity, vernacular weaved with formal
- **James Baldwin** — long balanced sentences, moral urgency in clause stacking, semicolons used as breath marks
- **Annie Dillard** — observational compression, present-tense immediacy, sentence as image
- **Marilynne Robinson** — theological cadence in plain words, long meditative sentences, Calvinist patience

## Essayists / analytical

- **Paul Graham** — thinking out loud, simple words, short sentences mixed with one long earned conclusion
- **Patrick McKenzie** — long discursive sentences with parenthetical asides, domain-specific precision, dry humor
- **Tim Urban** — extended metaphors, conversational asides, building up frameworks with named characters
- **Scott Alexander** — rationalist sectioning, exhaustive enumeration of possibilities, fair-witness analysis
- **Ben Thompson** — business-strategy decomposition, recurring framework names, lots of "this is why"
- **Tyler Cowen** — compressed, list-heavy, blogger-shorthand, range of references in one paragraph
- **Malcolm Gladwell** — narrative anchor → general principle → reversal, three-act essay structure
- **Adam Grant** — research-anchored, paired contrasts, gentle prescriptive framing

## Self-help / instructional / productivity

- **Mark Manson** — period-heavy clean prose, "Your X is Y" definitional moves, profane confidence, direct second-person
- **James Clear** — concept-coinage backbone, numbered enumeration, clean instructional cadence, named laws/frameworks
- **Ryan Holiday** — Stoic-instructional, drawing concepts from antiquity, clean delivery, repetition for emphasis
- **Tim Ferriss** — list-heavy, hack/protocol framing, second-person direct, capitalized concept names
- **Cal Newport** — academic-instructional, named frameworks, research-backed prescriptions, sober register
- **Greg McKeown** — one-idea-per-page rhythm, named principles, short paragraphs, prescriptive minimalism
- **Brené Brown** — vulnerability-as-rhetoric, personal anecdotes anchoring research, conversational warmth
- **Atomic Habits voice** — short-paragraph instruction, principle-then-example, mechanical-cause-and-effect language

## Polemicists / contrarians / philosophical-provocative

- **Nassim Nicholas Taleb** — concept-coinage (Black Swan, antifragile), aphoristic stabs, attacking IYI, ancient-thinker citations
- **Jordan Peterson** — lecture cadence, biological-evolutionary framing, religious overlay, definitional pivots
- **Bronze Age Pervert** — baroque Nietzschean, mock-archaic spelling, vitalist anti-modernity, ironic register
- **Curtis Yarvin** — reactionary historical, concept-naming as branding, sneering wit, long allusive sentences
- **Camille Paglia** — punchy contrarian, aesthetic-biological frame, dense allusion, no hedging
- **Bryan Caplan** — libertarian-economic, direct refutation, hypothetical thought experiments, plain professorial prose

## Tech / startup

- **Naval Ravikant** — aphoristic, tweet-shaped, concept-as-brand, distilled-wisdom cadence
- **Marc Andreessen** — manifesto-mode, accumulating short declarative lines, exhortation register
- **Sam Altman** — short, contrarian, "obvious in hindsight" framing, blog-post brevity
- **Peter Thiel** — paradox-as-thesis, philosophical-tech crossover, Strauss-influenced indirection
- **Joel Spolsky** — conversational tech-blog, anecdote-then-principle, signposting humor
- **Steve Yegge** — long discursive rants, programmer-culture inside jokes, accumulating digressions

## Narrative non-fiction / journalism

- **Michael Lewis** — character-first reporting, scene-as-argument, clean unobtrusive prose
- **Sebastian Junger** — documentary precision, anthropological framing, present-tense narrative
- **Jon Krakauer** — present-tense urgency, sensory immediacy, restraint in adjective use
- **John McPhee** — list-as-paragraph, structural patterning, long sentences with specific detail
- **Ta-Nehisi Coates** — meditative-historical, repetition as emphasis, address as form of argument
- **Tom Wolfe** — New Journalism, exclamation, capitalization, italics for sound, voice-jumping
- **Joan Didion (essays)** — see literary; her journalism has the same compressed precision

## Memoirists / personal voice

- **David Sedaris** — comic understatement, family scenes, deadpan one-liners as paragraph closers
- **Anne Lamott** — confessional warmth, self-deprecating humor, sentence fragments for emphasis
- **Anthony Bourdain** — profane confidence, food as window, baroque vocabulary mixed with kitchen-slang
- **Rick Bragg** — Southern oral cadence, specific-detail compression, sentence rhythms from speech
- **Mary Karr** — lyric memoir, line-break-tight sentences, Catholic-rural register

## Aphorists / brief-form

- **La Rochefoucauld** — epigrammatic, paired antithesis, cynical wit in one breath
- **Friedrich Nietzsche** — aphoristic, hammer-blows of declaration, philosophical provocation
- **E.M. Cioran** — pessimistic aphorism, paradoxical brevity, polished despair
- **Eric Hoffer** — longshoreman intellectual, declarative wisdom, sociological observation

## Academics-for-public

- **Steven Pinker** — cognitive-science precision, numbered argument, defending Enlightenment, ironic asides
- **Richard Dawkins** — precise zoological prose, metaphor as scaffolding, polemical clarity
- **Robert Sapolsky** — neuroendocrine framing, dense citation, jokes nested in dense paragraphs
- **Carl Sagan** — cosmic-scale lyric, science-as-wonder cadence, accessible majesty
- **Daniel Kahneman** — System 1 / System 2 framing, dispassionate exposition, named cognitive biases
- **Yuval Noah Harari** — sweeping historical synthesis, declarative simplification, "imagined orders" type concepts
- **Jared Diamond** — continent-scale comparison, geographic-determinist framing, list of factors enumeration

## Religious / philosophical (modern)

- **C.S. Lewis** — clear analogical prose, common-sense apologetics, "imagine that" framing
- **G.K. Chesterton** — paradox-as-rhetoric, joyful contrarian, accumulating images per sentence
- **Thomas Merton** — contemplative prose, slow paragraphs, Catholic-Buddhist hybrid

## Business / leadership

- **Peter Drucker** — dry analytical, principle-then-example, professorial calm
- **Andy Grove** — engineering-direct management, framework-naming, lean instructional
- **Ben Horowitz** — war-stories anchored to lessons, hip-hop epigraphs, conversational toughness

## Short-form / Twitter native

- **Visakan Veerasamy** — thread-shaped, hyperlinked references, generous tone, recursive callbacks
- **Hari Kondabolu / Twitter-essayist** — one-liner punch followed by longer unpacking, comic timing

## Genre-specific (modern poetic / lyric prose)

- **Ocean Vuong** — lyric memoir, line-conscious paragraphs, image-as-argument
- **Maggie Nelson** — theory-personal hybrid, numbered fragments, citations interleaved

## Reminder

These style notes describe **prose features only**. When matching the user's corpus against an author from this list, cite the author's prose feature — not their content or topic. If you find yourself matching "the user writes about topic X therefore resembles author Y," stop and re-do the match. The correct framing is "the user uses period-heavy declarative cadence with definitional pivots, which matches Peterson's lecture-cadence prose pattern."

# Fingerprints Catalog

> Exact presentation choices the author makes consistently. LLMs default to training-data defaults for all of these unless explicitly told the user's choice. So we measure each one and emit a one-liner.
> Use this when analyzing a user's corpus.

## How to use this catalog

For each fingerprint below:

1. Read the user's corpus.
2. Count the relevant variants.
3. Apply the decision rule (mostly ≥3 total observations, then a ratio threshold).
4. Emit a one-line fingerprint line if confidence ≥ medium. Skip if `n/a` (not enough signal).

The output goes into `voice/fingerprints.md` as a bullet list, each line in the format `<Label>: <value>`.

## Confidence rule (used by most binary fingerprints)

Total observations = `a + b` where `a` is the count of one variant and `b` the other.

- If total < 3 → **`n/a` (low confidence)**, skip emitting.
- If `a / total ≥ 0.85` → **`a` wins, high confidence**, emit.
- If `a / total ≤ 0.15` → **`b` wins, high confidence**, emit.
- If `0.70 ≤ a / total < 0.85` → **`a` wins, medium confidence**, emit.
- If `0.15 < a / total ≤ 0.30` → **`b` wins, medium confidence**, emit.
- Otherwise (0.30 < ratio < 0.70) → **`mixed` (low confidence)**, emit as "inconsistent" only if the user explicitly wants to see mixed signals; otherwise skip.

For the curious: this is asymmetric because we want either a clear choice (≥70%) or to skip. The 0.30-0.70 band is "the author isn't actually making a consistent choice" — emitting a fingerprint there would mislead.

## The eight fingerprints

### 1. Em-dash spacing

What to measure:
- `spaced` count = occurrences of `<whitespace>—<whitespace>` (e.g., `word — word`)
- `unspaced` count = occurrences of `<non-whitespace>—<non-whitespace>` (e.g., `word—word`)

Apply the binary confidence rule.

Output map:
- `spaced` → `Em-dash spacing: word — word`
- `unspaced` → `Em-dash spacing: word—word`
- `mixed` → `Em-dash spacing: inconsistent`

Note: if the user is on a NEVER em-dashes rule (didn't clear the punctuation hurdle), skip this fingerprint entirely.

### 2. Ellipsis style

What to measure:
- `three_dots` count = occurrences of `...` (three ASCII dots, not part of a longer run)
- `unicode` count = occurrences of `…` (single Unicode character)
- `spaced_dots` count = occurrences of `. . .` (dots separated by spaces)

Decision:
- If total < 3 → `n/a`, skip.
- If max variant / total ≥ 0.85 → emit the winning style.
- Otherwise → `mixed`.

Output map:
- `three_dots` → `Ellipsis style: ...`
- `unicode` → `Ellipsis style: …`
- `spaced_dots` → `Ellipsis style: . . .`
- `mixed` → `Ellipsis style: inconsistent`

### 3. Sentence-initial conjunction capitalization

What to measure:
- `caps` count = occurrences of `[.!?]<whitespace>(But|And|So|Or|Yet)\b` — i.e., starts a new sentence with capitalized conjunction
- `lower` count = occurrences of `[.!?]<whitespace>(but|and|so|or|yet)\b` — same but lowercase (unusual, only if author uses a stylistic comma-after-period thing)

Apply the binary confidence rule.

Output map:
- `caps` → `Sentence-initial "But/And/So": capitalized (". But")`
- `lower` → `Sentence-initial "But/And/So": lowercase (", but")`
- `mixed` → `Sentence-initial "But/And/So": inconsistent`

### 4. Oxford comma

What to measure (rough — accept some noise):
- `withOxford` = sequences matching `<word>, <word>(...), and <word>` or `<word>, <word>(...), or <word>`
- `withoutOxford` = sequences matching `<word>, <word>(...) and <word>` or `<word>, <word>(...) or <word>` (no comma before "and"/"or")

Apply the binary confidence rule.

Output map:
- `yes` → `Oxford comma: yes`
- `no` → `Oxford comma: no`
- `mixed` → `Oxford comma: inconsistent`

### 5. Quote style

What to measure:
- `straight` count = occurrences of `"`
- `curly` count = occurrences of `"` or `"`

Apply the binary confidence rule.

Output map:
- `straight` → `Quote style: straight "..."`
- `curly` → `Quote style: curly "..."`
- `mixed` → `Quote style: inconsistent`

### 6. Capitalization after colon

What to measure:
- `upper` count = occurrences of `: <Uppercase letter>`
- `lower` count = occurrences of `: <lowercase letter>`

Apply the binary confidence rule. Require total ≥ 3.

Output map:
- `upper` → `Capitalization after colon: upper`
- `lower` → `Capitalization after colon: lower`
- `mixed` → `Capitalization after colon: inconsistent`

### 7. Contractions

What to measure:
- `contracted` count = occurrences of `<word>'<s|re|ve|ll|d|t|m>` (e.g., `don't`, `I'm`, `we'll`)
- `expanded` count = occurrences of the literal phrases `do not`, `does not`, `did not`, `is not`, `are not`, `was not`, `were not`, `cannot`, `will not`, `would not`, `should not`, `could not`, `have not`, `has not`, `had not`, `I am`, `you are`, `we are`, `they are`, `it is`, `that is`, `there is`, `let us`

Apply the binary confidence rule.

Output map:
- `yes` → `Contractions: uses contractions`
- `no` → `Contractions: avoids contractions`
- `mixed` → `Contractions: mixed`

### 8. Paragraph length

What to measure:
- Split the corpus on double-newlines (`\n\n+`) to get paragraphs.
- For each paragraph, count sentences (split on `[.!?]<whitespace>`).
- Compute the average sentences-per-paragraph.

Require at least 3 paragraphs. Otherwise `n/a`, skip.

Decision:
- avg ≤ 2 → `short`, high confidence
- 2 < avg ≤ 4 → `medium`, high confidence
- avg ≥ 6 → `long`, high confidence
- 4 < avg < 6 → `mixed`, medium confidence

Output map:
- `short` → `Paragraph length: 1–2 sentences`
- `medium` → `Paragraph length: 3–4 sentences`
- `long` → `Paragraph length: 5+ sentences`
- `mixed` → `Paragraph length: varied`

## Output structure

The agent writes `voice/fingerprints.md` as:

```markdown
# Presentation Fingerprints

> Auto-generated from `voice/corpus/`. Exact presentation choices the user makes consistently.
> Match these in every generated response — LLMs default to training data otherwise.

- Em-dash spacing: word — word
- Oxford comma: yes
- Quote style: straight "..."
- Contractions: uses contractions
- Paragraph length: 3–4 sentences

## Manual Overrides

<!-- Override any auto-detected fingerprint here. These win over the auto-detected ones above. -->
<!-- Example: -->
<!-- - Em-dash spacing: never use em-dashes at all. -->
<!-- - Quote style: straight always. -->
```

The `## Manual Overrides` section MUST be preserved across regenerations.

# Authenticity Hurdle

> The core decision rule: when does a pattern in the user's corpus count as "their voice" versus "AI contamination"?

## The problem

If we just emit NEVER rules for everything in the AI tells catalog, we'll strip out words the user actually likes and uses. Example: Mark Manson regularly uses "crucial" and "vital." Banning those flattens his voice.

If we don't emit any rules unless the word appears zero times, we keep AI contamination. Example: an essay that the user originally drafted with ChatGPT help and then edited still contains stray "delves" and "valuable insights" — that's not their voice, that's leftover residue.

The hurdle resolves this: a pattern is "authentic" only if it appears at signature frequency. Below that → contamination, ban it.

## The thresholds

| Category | Min count | Min rate per 1000 words | Why |
|----------|-----------|-------------------------|-----|
| `word` | 2 | 0.15 | Single words are noise-prone. Need at least 2 and a non-trivial rate. |
| `transition` | 2 | 0.15 | Same as words. |
| `phrase` | 2 | 0.05 | Phrases are more distinctive — a lower rate still signals deliberate use. |
| `punctuation` | 15 | 1.5 | Punctuation is denser than diction. A few em-dashes is normal; signature use means many. |

## The decision

For each AI tell, given `count` (how many times it appears in the corpus) and `words` (total word count of the corpus):

```
rate_per_1k = (count / words) * 1000
passes = count >= min_count AND rate_per_1k >= min_rate_per_1k
```

If `passes` → **preserve** (no NEVER rule). The user genuinely uses this.

If not `passes` → **forbid** (emit NEVER rule). This includes:
- Items that never appear (forbid because training data will push them back in)
- Items that appear once or twice but below the rate threshold (forbid because it's likely contamination, not signature)

## Worked examples

**Word "delve" in a 2000-word corpus:**
- count = 0 → fails (count < 2) → **forbid**: `NEVER "delve".`
- count = 1 → fails (count < 2) → **forbid**: `NEVER "delve".` _(below hurdle — flagged in status.md as likely contamination)_
- count = 2, rate = 1.0/1k → passes (count ≥ 2, rate ≥ 0.15) → **preserve**, no rule.

**Em-dashes in a 5000-word corpus:**
- count = 8, rate = 1.6/1k → fails (count < 15) → **forbid**: `NEVER "em-dashes".`
- count = 20, rate = 4.0/1k → passes → **preserve**, no rule.

The em-dash hurdle is intentionally hard. Most human writers don't clear it. The few who do (people with serious published-prose backgrounds) get to keep their em-dashes.

## What this implies about tier progression

The skill's tier system tracks corpus word count:
- Tier 1 (300–1k words): too small to clear most hurdles. NEVER rules emit aggressively.
- Tier 2 (1k–5k): some words start to pass. Most phrases still fail (they need ≥2 instances).
- Tier 3 (5k–20k): phrases and many words can pass. Em-dash hurdle still hard.
- Tier 4 (20k+): em-dash hurdle achievable. Profile is AV-grade.

The hurdle stays the same across tiers. What changes is that more corpus means more chances for the user's signature patterns to clear it.

## Status reporting

For items that appear in the corpus but fail the hurdle (the "below_hurdle" set), flag them in `voice/status.md` under a "Below-Hurdle Detections" section. Format:

```
- `delve` — appears 1x, rate 0.5/1k
- `however` — appears 1x, rate 0.5/1k
```

This is informational — the user can see what got flagged as contamination. It helps them notice "oh, I have a habit of letting AI drafts through" or alternatively "wait, I actually do use that word, let me add it to manual preserves."

## Edge cases

- **Count = 0**: always forbid. Don't list in below_hurdle (because it isn't "present in corpus").
- **Tied counts in fingerprint binary**: if `a == b` (exactly equal), treat as `mixed`.
- **Total < 3 in fingerprint**: skip entirely — not enough signal.
- **Negative numbers, NaN, infinity**: shouldn't happen, but guard with `count = max(0, count)` if you implement defensively.

# Post-Write Audit

> Distribution-level statistical checks the orchestrator runs against the minion's returned prose. Sits between step 6 (NEVER scan) and step 7 (integration) of the Apply Protocol. Catches statistical fingerprints the minion's prompt can't reasonably prevent without cognitively overloading the writing pass.

## When this runs

After step 6 (NEVER-violations scan + brief-error patching), before step 7 (integration). The orchestrator reads this file, applies each check to the minion's output, and surgically rewrites the smallest span that brings the failing metric back into range.

## Why this layer exists

The minion writes prose. The orchestrator polices distribution and lexicon. Anything mechanically detectable after the fact lives here, not in the writing-pass prompt — the minion's cognitive budget should go to channeling the anchor and hitting the commitments, not tracking 60 micro-bans.

Two enforcement points still exist for the bans the minion DOES need to see (contrastive negation, sentence-opener repetition, em-dashes, etc.) — those live in `voice/never-rules.md` and get scanned at step 6. This audit is for the slop that's cheaper to scrub than to prevent.

## Remediation principle

For each failing check, rewrite the **smallest local span** that fixes the metric. Do not regenerate. Do not reach for stylistic improvement. The minion's voice IS the result — the audit only nudges the statistics.

If a failing span is load-bearing (a specific image, a coined term, a structural beat the brief demanded), leave it. Audit findings are advisory at the boundary case. The minion's intent wins ties.

Aim for the lightest touch: 5-10 small substitutions across a typical draft brings rates back in line. Heavier rewrites mean the audit is being misused.

## Distribution checks

### 1. Sentence-opener repetition

**What to measure:** walk the output sentence by sentence. For each window of 3 consecutive sentences, check whether all three start with the same first word.

**Threshold:** flag if >30% of windows trigger.

**Why this number:** human writing sits at ~17% (DFT 2026 — mostly from intentional list structures like "How does X?... How does Y?... How does Z?"). SFT models at T=0.7 hit 53.3%. The 30% line cleanly separates human from AI.

**Action:** locate the offending windows. For each, rewrite the second OR third sentence to start with a different word. If the window forms an intentional list, leave it — list structure is the human use case the 17% baseline reflects.

### 2. Sentence-initial "The" frequency

**What to measure:** percentage of sentences that begin with the word "The."

**Threshold:** flag if >15% of all sentences.

**Why this number:** "The" at sentence start is over-used by ~90% in SFT output vs human writing (DFT 2026, 14B SFT model). The +90% inflation puts AI rates well above the natural human range.

**Action:** locate sentences starting with "The." Rewrite a portion to start with a different determiner ("A", "An", "These"), a pronoun, a prepositional phrase, or a different subject. Five to seven swaps across a typical paragraph is usually sufficient.

### 3. Function-word over-use

**What to read for:** the AI distribution-distance signal lives mostly in function words, not fancy diction. Top-10 tokens account for 87.2% of L2 distribution distance in SFT output (DFT 2026). Watch for:

| Token | SFT inflation vs human |
|---|---|
| `is` | +44% |
| `was` | +49% |
| `are` | +31% |
| `that` | +25% |
| `a` | +15% |
| `to` | +11% |
| `.` (period) | +19% |

**Heuristic check (no exact threshold):** scan the draft for clusters of short copular sentences ("X is Y. Z is W. P is Q.") and high period density (many short sentences in a row). Both are signatures of function-word inflation.

**Action:** when noticed, merge two short copular sentences into one with a participial or relative clause; vary sentence structure to use action verbs instead of "is/was"; combine short sentences to drop period count. Three to five rewrites across a paragraph usually levels the distribution.

### 4. Sentence-length variance

**What to measure:** compute standard deviation of sentence length (in words) across the output. If the user has a `voice/stats.md`, compare to the user's own σ. Otherwise compare to baseline σ ≥ 8 words.

**Threshold:** flag if σ < 6 words (low variance — uniform sentence length is an AI signature).

**Action:** locate runs of similar-length sentences. Merge two short ones into a longer compound, or split a medium one. The goal is to restore length variance, not hit a specific number.

## Lexical watch list

Mechanical word-level scrubs. The minion doesn't see these — the audit handles them on the way out.

### GPT-5 specific over-used tokens (DFT 2026)

When the minion is a GPT-5-class model, these tokens are inflated vs human writing. Scan for them:

| Token | Inflation vs human | Human baseline |
|---|---|---|
| `corridors` | +45.2% | 0.1% |
| `norms` | +43.1% | 0.1% |
| `align` / `aligns` / `alignment` | +36.0% | 0.2% |
| `metrics` | +27.2% | 0.2% |
| `engagement` | +26.5% | 0.2% |
| `targeted` | +5.1% | 1.6% |
| `identity` | +5.0% | 1.0% |
| `trust` | +4.9% | 1.2% |

**Action:** swap to a context-appropriate alternative when the word appears in surplus (3+ uses in a short piece, OR any use in a context where the word feels generic). If the user's corpus contains the word at signature frequency (in `voice/stats.md` or `voice/never-rules.md` exempts), leave it — they own that word.

### Named-character defaults

AI defaults to specific generated names in fiction. Known examples:

- `Elara Voss` — documented in OpenAI's "goblin problem"
- Add new defaults as documented.

**Action:** if found in fiction output without explicit user specification, rename to something contextually appropriate or to a name the user has used in their corpus.

## Source

Distribution thresholds and over-use rates from "Fixing LLM Writing with Distribution Fine-Tuning," Rosmine 2026 (https://rosmine.ai/2026/05/18/fixing-llm-writing-with-distribution-fine-tuning/). Token over-use rates measured against 14B SFT vs human fineweb baseline. Sentence-opener repetition methodology: percent of texts containing 3+ consecutive sentences starting with the same first word.

This file is a living checklist. New research that surfaces measurable thresholds for AI-vs-human writing belongs here, not in `voice/never-rules.md` — the writing pass stays lean.

# Analysis Protocol

Regenerates the voice files from the corpus. Run any time the corpus changes (new samples added, samples removed, samples revised). Loaded only when triggered — not in context during normal writing sessions.

## Protocol

1. **Read inputs.** Concatenate every file in `voice/corpus/` (strip frontmatter). Count words. Read `catalog/ai-tells.md`, `catalog/fingerprints.md`, `catalog/hurdle.md`.

2. **Compute deterministic tally** (best effort — counts may drift ±1 on long corpora):
   - **Sentence distribution**: split on `[.!?]\s`, compute short/medium/long/very-long percentages, average length. Set `short_max` (25th-pct, clamped [6,12]) and `long_min` (75th-pct, clamped [18,28]). Do NOT emit a sentence-length cap in the apply directive — the corpus distribution carries the right ceiling and an arbitrary cap suppresses signature long sentences.
   - **Punctuation density per 1k words** for em/en dash, colon, semicolon, question, exclamation, ellipsis, paren, bracket, straight/curly quotes. Categorize as `never` / `rare` / `low` / `strong`.
   - **AI-tell tally**: count each item from `catalog/ai-tells.md`. Apply hurdle from `catalog/hurdle.md`: passes hurdle → preserve; fails → emit NEVER rule; below-hurdle but present → log to `below_hurdle`.
   - **Fingerprints**: apply each detector from `catalog/fingerprints.md` with its decision rule.

3. **Determine tier** by word count: <300 = 0 Empty; 300-999 = 1 Anchor; 1000-4999 = 2 Preliminary; 5000-19999 = 3 Full Coverage; ≥20000 = 4 AV-Grade. See `docs/tiers.md` for what unlocks at each tier.

4. **Write `voice/stats.md`** — corpus stats, sentence distribution table, punctuation density table.

5. **Write `voice/never-rules.md`** — preserve `## Manual Additions` section verbatim (anchored to start-of-line; the literal also appears in the intro blockquote — naive search will mis-grab it).

6. **Write `voice/fingerprints.md`** — preserve `## Manual Overrides` section, same caution.

7. **Write `voice/status.md`** — tier, words, active features, locked features, next milestone, file list, below-hurdle detections.

8. **Report** — new tier, what changed in NEVER rules, what's locked next.

For corpora >10k words, count in passes (words → phrases → transitions) rather than tracking 60 counters at once.

## Adding Samples Later

User says "add this to my voice profile" or pastes new writing. Append to next `voice/corpus/sample-NNN.md`, re-run Analysis Protocol, report tier change if any.

# Anchor Iteration

Final-polish minion. Channels the user's voice anchors as a panel; iterates critique → rewrite → re-score until 90/100. Single minion conversation, visible iteration history. Mandatory anti-ai cleanup follow-up.

Specialization of `/polish`'s pattern for writers-voice: channels the user's specific voice anchors (dynamically loaded from `voice/anchor.md` or `voice/anchor-<context>.md`), not generic advertising practitioners.

Replaced the prior single-pass Anchor Critique tool. Single-pass scoring is now just "stop after iteration 1" of Anchor Iteration — same architecture, parameter difference.

## Why this works

AI cannot judge beats subjectively from generic prompts — no dopamine system to consult. But channeled anchors carry beat-judgment encoded in their training-data representations. Channeling Peterson reading prose surfaces Peterson's beat-trained sensibility. Multiplied across the panel, the collective weighted score reflects how the prose lands across the writer's actual voice ambition.

This is the ONLY AI critic tool that can do beat-level judgment, and it works only because the anchors are real humans whose dopamine-trained sensibilities are encoded in training data. Judgment is collective and writer-specific.

## Architecture: one minion, visible iteration

ONE minion conversation. The minion runs the entire iteration loop internally with full visible history — each iteration's anchor critiques are part of the minion's context for the next iteration. Anchors see how their previous critiques were addressed (or weren't), which sharpens subsequent critiques.

Mirrors `/polish` exactly. NOT extracted-and-rerun-cold per iteration. The loop has memory.

Optional fallback: if visible iteration converges on a local optimum (rewriter anchored to original framing through visibility), retry with blind iteration (no prior history shown across iterations). Default = visible.

## Inputs (only)

- The prose to polish
- The voice anchor blend (dynamically loaded from `voice/anchor.md` or `voice/anchor-<context>.md`)

That's the complete input. No commitments. No beat sheet. No project context. Anchors read the prose AS-IS, like a reader encountering it cold.

**Why no context:** anchors must judge the prose AS IT LANDS, not as it was intended to land. Briefing them on the project would have them judge against the brief, not against the prose. Cold reading is the point.

## Personas (dynamic, inferred)

Pulled from `voice/anchor.md` (or context-specific variant). Each anchor listed with a blend percentage that serves as both voice influence weight and panel vote weight.

The minion infers each anchor's persona from training data — named writers are known entities to opus. System prompt does NOT enumerate per-persona profiles. `/polish` works the same way ("top 10 advertising practitioners" — no profiles needed; opus knows Hopkins, Ogilvy, Sugarman, etc.).

Example book-project anchor file:

```
- 26% Jordan Peterson
- 22% Robert Sapolsky
- 20% Nassim Taleb
- 18% Bryan Caplan
- 14% Naval Ravikant
```

Minion channels each with their characteristic sensibility — Peterson for moral weight and structural rigor, Sapolsky for biological grounding and dry mechanism, Taleb for skin-in-the-game and aphoristic hardness, Caplan for clear thesis with evidence, Naval for aphoristic screenshot-worthy compression.

## Process (per iteration)

1. Each anchor reads the current prose AS-IS, with their characteristic sensibility
2. Each anchor produces:
   - SCORE 0-100 (honest read of how this lands for them)
   - TOP CRITIQUE (ONE thing they would cut, sharpen, or restructure)
   - STRENGTH (what's working that must be preserved)
3. Compute collective weighted score using anchor blend percentages
4. If collective ≥ 90/100: STOP. Mark as FINAL ITERATION. Return converged prose.
5. If collective < 90/100: synthesize panel's critiques into a FULL REWRITE of the prose (not a sentence-level patch). Preserve named strengths.
6. Begin next iteration with rewritten prose. Each iteration's anchors see all previous iterations and critiques in their context.

ITERATION CAP: 6 iterations. If still <90 after 6, return highest-scoring iteration with note about non-convergence.

## Output format

Per iteration:

```
====== ITERATION N ======

PROSE (current state):
[the prose being read this iteration]

ANCHOR READINGS:

{Anchor Name} ({weight}%): SCORE: X/100
  TOP CRITIQUE: ...
  STRENGTH: ...

[repeat per anchor]

COLLECTIVE WEIGHTED SCORE: X/100

[if < 90:]
SYNTHESIS — what the rewrite must address:
- ...

REWRITE:
[next iteration's prose, full text]

[if ≥ 90:]
CONVERGED. Final prose ready below.
```

After convergence (or cap):

```
====== FINAL ======
Iterations: N
Final score: X/100
Convergence: YES / NO

FINAL PROSE:
[the polished prose, full text]
```

## Mandatory anti-ai follow-up

Anchor iteration runs no-context. NEVER rules and presentation fingerprints are not in scope during iteration. Rewriter will introduce AI tells the original prose may have avoided.

After convergence, editor MUST run an anti-ai cleanup pass against `voice/never-rules.md` and `voice/fingerprints.md`. Common scrubs:

- em-dashes → commas, periods, or restructured sentences
- semicolons → "and" or new sentences
- contrastive negation patterns → direct positive statements
- banned diction → plain equivalents
- inserted parenthetical em-dashes → restructure

TWO-STEP pattern. Iteration THEN anti-ai. Two passes do different jobs and should not be conflated. Skipping the anti-ai pass ships AI fingerprints into the published prose.

## When to fire

- Final polish of a beat or section before publishing
- After integration of multi-minion drafts is complete and the Blinder Audit is clean
- When prose is content-finished and needs to land at ship-level voice quality

Do NOT fire on:

- Rough first drafts (commitments may still be evolving — polish wastes effort)
- Sections under structural revision (rewrite the commitments first, then polish the result)
- Single-paragraph fragments (panel needs prose to evaluate; isolation produces weak critiques)

## Editor's role

1. Identifies a beat ready for final polish
2. Fires Anchor Iteration with the prose + dynamically loaded anchor blend
3. Receives the converged output
4. Runs the mandatory anti-ai cleanup pass
5. Posts the result for review or comparison

The editor does NOT:
- Inject project context into the iteration (preserves cold-reader purity)
- Stop the iteration early (let convergence happen — the loop is the point)
- Re-judge the panel's collective decisions (panel's authority is the entire point of the tool)

## Failure modes

- **Sycophantic clustering**: anchors give 85+ uniformly. Prompt explicitly bans default-middle scoring and names what 90/60/40 means for each anchor (90 = anchor would actually quote / share; 60 = competent but forgettable; 40 = anchor would put it down).
- **Persona drift**: anchors all sound like generic helpful AI. Combat by instructing channel-faithfully — each anchor should sound like the actual writer, hostile to AI flattening.
- **Iteration plateau**: scores stop rising after iteration 3-4. Panel has done what it can. Ship the highest iteration even if below 90, or fall back to blind iteration.
- **Manufactured content**: rewriter may invent details to address critiques (fabricated autobiographical claims, invented statistics, manufactured anecdotes). Editor MUST scan output for invented content during anti-ai pass and verify or cut as appropriate.

## Model

**Opus required.** Sonnet drifts to default-helpful behavior, scores uniformly high, produces weak rewrites that don't actually address critiques. Opus channels personas with discipline and produces rewrites worth re-scoring.

## Cost

Single conversation, multiple turns. Per iteration: ~3-5k input tokens (prose + previous iteration history) + ~3-5k output (critiques + rewrite). Three iterations ≈ 30k total tokens. Acceptable for chapter-scale work.

For short pieces (tweets, single paragraphs), Anchor Iteration is overkill. Use `/anti-ai` alone or a single Apply call with strong commitments.

## Validation

Tested 2026-05-18 on a 1000-word chapter beat. Three iterations:

- **Iteration 1**: 77.20/100. Panel flagged: thesis buried, mechanism overclaim, no skin-in-the-game, no screenshot-worthy lines.
- **Iteration 2**: 86.32/100. Rewrite added thesis paragraph + personal admission + mechanism hedge + sharpened closing teaser.
- **Iteration 3**: 91.04/100. Strengthened personal admission, fixed determinist phrasing, added specific enemies, gave load-bearing line its own paragraph. Converged.

Notable failure modes observed:

- Rewriter manufactured an autobiographical detail (Iteration 2 invented a personal admission to satisfy a skin-in-the-game critique). Editor caught and flagged for verification during anti-ai pass.
- Iteration introduced em-dashes across 8 paragraphs and semicolons in the opening (original prose used "and"). Mandatory anti-ai pass scrubbed all of them.

The tool delivered ship-quality prose in 3 iterations. The post-iteration anti-ai pass was non-optional.

# Apply Protocol — Deep Reference

Loaded when scoping a brief (Apply step 4) or running cross-section coherence review (Apply step 8). Not in context for routing or other turns.

## Step 4 — Writing the TASK brief

Promoted to `SKILL.md` Apply Protocol step 4. The load-bearing rules (commitments-only default, never meta-references, COMMITMENTS as quasi-verbatim, read-prior-integrated-sections, preservation scope, cadence prescription) live there. This doc holds the edge-case templates and minion taxonomy below.

## Writing minion taxonomy

| Minion | Scope | Input | Output | When fires |
|---|---|---|---|---|
| **Apply** | Generative writing | Commitments + voice + context SUMMARIES (no source prose) | Fresh prose | Initial drafts. Also small reframes when audit prescribes a structural fix (audit's prescription becomes the commitment). |
| **Rewrite** | Generative writing against updated commitments WITH context awareness | Updated commitments + voice + context layers (summary of preceding + key-term glossary + adjacent seam paragraphs as orientation-only) + cadence prescription. NO source prose for content being written. | Fresh prose | Beat Map commitments changed; surrounding prose is acceptable and should be preserved. Apply brief + context awareness layers. |

Plus **Blinder Audit** (Step 8b — paragraph-level substance duplication, critic only) and **Anchor Iteration** (Step 10 — final polish, channels voice anchors as panel, iterates to 90/100; see `anchor-iteration.md`).

### Editor's classification job after Audit fires

Each finding routes to one of three paths:
- **Editor direct (no minion)** — cuts, swaps, reorders, sentence-level surgical patches. Per FIRM RULE 1 carve-out.
- **Rewrite Minion** — paragraph's SUBSTANCE needs to change (structurally redundant with another, or mission needs to shift). Compression alone produces thin output.
- **Apply Minion (small)** — audit prescribes a specific reframe of a small span and new prose is needed (e.g., "reframe opening sentence to acknowledge pivot from X to Y"). Audit's prescription becomes the commitment.

Classification rule: **before flagging anything for a minion, ask "is this a SHAPE problem or a SUBSTANCE problem?"** Different work needed → Rewrite or Cut. New prose into a small span → small Apply. Surgical word/sentence work on existing prose → editor direct.

## Rewrite Minion

The Rewrite Minion IS the Apply Minion fired with context awareness. The brief shape preserves Apply's no-source-prose-ceiling property (minion brings its own moves) while adding context layers needed to flow into surrounding prose.

### Validated brief shape

```
[Voice profile: anchor, NEVER rules, fingerprints, stats, coined terms, examples]

---

TASK:

PROJECT: [1-paragraph context — what the doc is, what the section does]

CONTEXT — what's already established in the surrounding prose:
- [Summary of preceding section as bullets — what was named, what was claimed, what threads are live. NOT raw prose. 5-10 bullets.]

KEY TERMS already named (available for callback if useful):
[Glossary list — coined terms, named concepts, distinctive phrasings the prose has established]

IMMEDIATELY PRECEDING PARAGRAPH (for seam continuity ONLY; do NOT mirror its cadence):
[Full paragraph verbatim — 1 only]

IMMEDIATELY FOLLOWING PARAGRAPH (for seam continuity ONLY; do NOT mirror its cadence):
[Full paragraph verbatim — 1 only]

COMMITMENT(S) — what your new content must land in the reader:
[Outcome statement(s) per beat]

CADENCE PRESCRIPTION (per paragraph):
[Explicit rhythm direction — open with X, build with Y, close with Z. Mandatory.]

LENGTH: [target word count]

Return prose only. No commentary. No headers. No beat labels.
```

The two seam paragraphs are full prose (minion needs flow continuity at the join), flagged explicitly as orientation-only. This avoids the source-prose ceiling because the task is "write fresh content that flows from / into these" rather than "match this cadence." **Framing matters more than presence.**

### Why these elements

- FULL adjacent seam paragraphs flagged "orientation only — do NOT mirror cadence" — flow continuity without ceiling
- SUMMARIES of preceding section (bullets, not raw prose) + key-term glossary — document-scale awareness
- OMITTING source prose for content being written — preserves no-ceiling property
- INCLUDING explicit cadence prescription per paragraph — mandatory; short calls need cadence MORE, not less

### Scope

Works at any scope: paragraph-level (1 new paragraph), section-level (3-5 paragraphs), beat-level (full beat regeneration). Pick by what changed in commitments — one beat's outcome → regenerate one paragraph; chapter-arc beat's whole outcome shape → regenerate the whole beat; multiple beats → batch per chapter-arc beat.

### Preserving specific lines from existing prose

List them as MUST-APPEAR-VERBATIM in commitments (Apply's "selective lifts" mode). Rare — usually the writer's training data brings stronger lines than the existing prose anyway.

### Audit follow-up MANDATORY

Rewrite outputs are subject to the same minion-blinder problem as first-pass Apply. The minion only sees its slice — seam paragraphs + summary bullets, not every paragraph in adjacent sections. It can introduce repetitions with paragraphs it didn't see.

**After any Rewrite Minion call, fire the Audit Minion (Step 8b) on the integrated document.** No exceptions. Empirical case: first Rewrite test (B1 close + B2 opening) produced strong individual prose AND 8 audit findings — most notably a visual-roster repetition across 4 paragraphs the rewriter never saw.

### Don't conflate rewrites with violation patches

Violation patches (FIRM RULE 1 carve-out) are 1-3 sentence local fixes to NEVER violations or brief errors WITHIN otherwise-acceptable prose. Editor work, small scope, surgical. Rewrites here mean re-running the minion against UPDATED commitments — different scenario, different scope decision.

## Step 8 — Cross-section coherence review

After integrating multiple minion outputs, scan for what individual runs cannot see:

- **Cadence repetition** across sections (same opens, same closes, same paragraph counts)
- **Recurring metaphors / phrases** across sections
- **Structural sameness** (every section ends with 4-layer enumeration, every section opens with 3 shorts)
- **Coined term overuse** — coined terms get injected into every minion call as MUST-PRESERVE, producing document-scale repetition (e.g., "territory" every 3 paragraphs). Track heavy-use terms; for subsequent minions, omit heavy-use terms from coined-terms injection OR add "use sparingly" instruction

Fix options: re-spawn with varied prescription, surgical post-edit (combine/split/swap), vary commitments + coined-terms per section at brief-assembly time, or accept for low-stakes drafts. The editor owns document-scale coherence; minions are responsible only for section-scale quality.

## Step 8b — Blinder Audit Minion (mandatory post-integration)

### Purpose

Minions write with SURGICAL context — just their slice of the doc — to avoid source-prose ceiling and context pollution. The trade-off: minion A doesn't know what minion B wrote. They can independently produce paragraphs whose ENTIRE SUBSTANCE mirrors another paragraph's entire substance. The **blinder problem**.

The Blinder Audit Minion has ONE job: find pairs of paragraphs whose whole substance closely mirrors each other.

### Operational test

The audit must reduce to an OBJECTIVE pattern-match — not a subjective claim about reader experience. The working question:

**"Summarize each paragraph in one sentence. Are two paragraphs' summaries effectively the same?"**

Or sharper: **"Could I cut this paragraph entirely and lose only redundancy?"** Answerable by reading; neither requires reader-experience judgment.

Empirically validated: the visual-roster case (4 paragraphs each enumerating the same 5 species in different framings) was a real blinder hit. Every other category tested at sentence-level, transition-level, or image-anchor level was a false positive — either intentional craft (image-anchoring, scaffolding, callbacks) or AI baseline patterns the reader doesn't notice.

### When to fire (mandatory)

- After integrating ≥2 minion outputs into one document
- After any rewrite cycle touching multiple sections
- Before showing the integrated document to the user

Skip when: single-minion / single-section work; surgical violation patches only.

### What to scan for (ONE category)

**PARAGRAPH-LEVEL SUBSTANCE DUPLICATION** — two paragraphs whose entire substance closely mirrors each other.

Diagnostic for any candidate pair:
1. Summarize paragraph A in one sentence. ("This paragraph does X.")
2. Summarize paragraph B in one sentence. ("This paragraph does Y.")
3. If X and Y are effectively the same job done with different words — finding.
4. If X and Y are different work — even if paragraphs share images, lexical phrases, or thematic threads — NOT a finding.

Cut test: could you delete one paragraph and lose only redundancy (no unique substance, no unique image-anchor, no unique scaffolding move)? Yes → real blinder. Cutting would lose something distinct → NOT a blinder regardless of surface similarity.

### What NOT to scan for (explicit exclusions — these produce false positives)

The audit must NOT flag any of these, even when surface similarity exists:

- **Sentence-level overlap across paragraph boundaries** (P4's closing shares thematic thread with P5's opening). Bridge/transition work. Cutting disconnects.
- **Image-anchoring across paragraphs** (the same image used in 2-3 paragraphs to thread a concept). Craft. The savanna/lion appearing in opener + body + closing is intentional thread.
- **Callbacks** (an image returning later as frame device or recognition moment). Craft.
- **Scaffolding repeats** (a paragraph opening with brief re-statement of previous paragraph's premise to set up its own new move). Premise-restatement is structure-promise, not duplication.
- **Sentence-level lexical/thematic overlap of any kind** unless part of WHOLE-PARAGRAPH substance duplication. Sharing the word "engine" or "same biology" is sentence-level — not a finding.
- **Structural sameness** (parallel cadence, repeated openers, declarative thesis + builds + aphorism). Readers don't notice; AI baseline behavior addressed by /anti-ai.
- **Cadence repetition** (parallel rhythm runs in adjacent paragraphs). Same reason.
- **Coined-term recurrence**. Coined terms are SUPPOSED to recur as identity markers.
- **Single-paragraph internal repetition** (triple anaphora within ONE paragraph). Craft, and the audit scans across paragraphs not within.

### Expected hit rate

Paragraph-level substance duplication is rare in semi-competent writing. On well-written beats the audit should typically return **NO BLINDER ERRORS FOUND** — the correct output, not a list of stretched findings to justify the call.

The audit fires as backstop on every integration. It catches gross duplications (visual-roster case, two paragraphs accidentally doing the same teaching beat). Most of the time it correctly returns zero.

### Report format

```
Finding #N
- LOCATION: paragraph references (e.g., "B2 P4 + B2 P5")
- SUMMARY A: one sentence describing what paragraph A does
- SUMMARY B: one sentence describing what paragraph B does
- WHY THE SUMMARIES MATCH: one sentence showing the substance duplication
- CUT TEST: could one paragraph be deleted entirely with only redundancy lost? (If no, the finding is invalid; do not include.)
- SEVERITY: moderate / major (paragraph-level duplication does not produce "minor" findings)
- SUGGESTED FIX: cut paragraph A / cut paragraph B / merge to single paragraph
```

If nothing meets the bar, return exactly: **"NO BLINDER ERRORS FOUND"** — that string, nothing else. Do NOT pad with sentence-level observations, structural notes, or polishing suggestions.

### Editor's response

For each finding:
- Apply the cut test independently. Does deleting one paragraph lose only redundancy?
- Yes → **cut** (editor territory, FIRM RULE 1 carve-out). Delete weaker paragraph via write_to_pad. If both are strong but redundant, merge unique fragments via Rewrite Minion.
- No → audit got it wrong. Defer.

Audit minion does NOT patch. Only reports.

Classification failure modes:
1. Trusting a finding without applying the cut test. If both paragraphs carry distinct substance, the audit was matching surface features. Defer.
2. Routing a paragraph-level cut to a Rewrite Minion when it should just be a cut. If two paragraphs do the same work, deleting one is the cleanest move.

### Brief shape (template)

Index every paragraph in the doc (e.g., "B1 P1: ...", "B2 P7: ...") so the audit can reference cleanly. Pass the full indexed document + the ONE category + the explicit exclusions + the output format. No voice profile needed (this isn't writing prose). Use opus.

# Context Hygiene

Reset context before applying voice to fresh writing — voice profiles fight against active conversation context and lose. Anchor blends, NEVER rules, and first-token cadence all get out-pulled by whatever prose dominates the live session.

## Two situations

| Situation | Practice |
| --- | --- |
| First piece of fresh writing in this session | **Reset.** Start a fresh session. Apply Protocol loads voice files cold. |
| Iteration on already-voice-applied writing (review → revise → review) | **Stay.** The context IS the voice you locked in. |

## When to surface the prompt

Surface only when ALL THREE hit:

1. Voice profile is set up at Tier 1+
2. Request is fresh writing, not iteration
3. Session has substantial prior context unrelated to the writing task

Skip for brand-new sessions or when prior context IS the writing-task setup.

## Prompt to surface

> Voice profile is set up at **Tier N**. Context here is polluted with **<one-line summary>**, which will pull output toward that register instead of the locked voice.
>
> For best output, start a fresh session and run:
>
> ```
> /writers-voice
> <then ask for your writing task>
> ```
>
> Or tell me **"write here anyway"** and I'll proceed with the active context.

# Setup, Anchor Protocol, Multi-Register

Loaded on first run, when generating a new anchor, or when splitting a corpus by register. Not in context during normal writing sessions.

## Setup Flow

If `voice/anchor.md` doesn't exist or is empty, walk the user through setup:

1. **Get the anchor.** Derivation is **fully local** — your own agent analyzes the writing, nothing leaves the machine, no service cost. Run the **Anchor Protocol** below to generate `voice/anchor.md` directly from the corpus on disk. Launch it as a sub-agent (see "Launching the anchor as a sub-agent") so the full stylometry rubric never pollutes the main session. Works offline. If the user has no corpus yet, ask them to seed 2-5 paragraphs (step 2) first — there is no hosted alternative.

2. **Seed the corpus** — ask for 2-5 paragraphs, write each to `voice/corpus/sample-NNN.md` with `added: YYYY-MM-DD` frontmatter.
3. **Run Analysis Protocol** (see `docs/analysis.md`) — populates `stats.md`, `never-rules.md`, `fingerprints.md`, `status.md`.
4. **Optional: curate examples** — ask the user for 3-5 most-representative paragraphs, write to `voice/examples.md`.
5. **Optional: populate coined terms** — ask the user for any coined terms / proper-noun concepts they want preserved verbatim, write to `voice/coined-terms.md` as a bare bullet list.
6. **Report status** — read `voice/status.md` and tell the user their tier + what unlocks next.

## Anchor Protocol (fully in-agent)

Generates `voice/anchor.md` (lean) and `voice/anchor-analysis.md` (rich).

1. Confirm corpus has ≥300 words. Below 300, ask the user to add a few more samples before anchoring.
2. Read `voice/stats.md`. If missing, run **Analysis Protocol** first.
3. Read `catalog/anchor-prompt.md` (full stylometry rubric) and `catalog/author-hints.md` (curated training-data authors with prose features).
4. **Set aside conversational context.** Score the corpus on prose mechanics only — never on themes/topics.
5. **Per-sample register analysis.** For each sample, record word count, address mode, register, signature moves. Flag samples >25% volume. Cluster by register; if 2+ distinct registers appear, flag as multi-register corpus.
6. **Score with register-aware feature validation.** Apply the 8 dimensions from `catalog/anchor-prompt.md`. Match against author hints. Assign weights summing to 100. For each cited feature, verify ≥40% sample appearance OR ≥40% volume (if neither, drop the feature; if it was the strongest evidence, drop the author).
7. **Self-criticism pass.** Strip any thematic reasoning. Set `confidence` and `any_thematic_reasoning` flags.
8. **Write `voice/anchor.md`** — JUST the lean `- N% Author` lines. No headers, no sub-bullets.
9. **Write `voice/anchor-analysis.md`** — per-author features, per-sample table, register diversity, self-check, refresh notes. Human-facing only.
10. **If multi-register corpus detected**, recommend a Multi-Register Split (see below).
11. Report blend + confidence + caveats to user.

## Launching the anchor as a sub-agent

The Anchor Protocol loads a large stylometry rubric (`catalog/anchor-prompt.md`),
the author-hints catalog, and the full corpus. Running it inline floods the main
session with analysis context the user never needs to see. **Launch it as a
sub-agent instead** — the sub-agent does the heavy reading and writes the files;
the main session gets back only a short summary.

Use the Agent/Task tool (general-purpose) with a self-contained prompt. The
sub-agent has no memory of this conversation, so the prompt must name every file
by absolute path:

> Generate a writer's-voice anchor, entirely locally. Do NOT call any network
> service or API — analyze with your own reasoning only.
> 1. Read the stylometry rubric at `<skill>/catalog/anchor-prompt.md` and the
>    author hints at `<skill>/catalog/author-hints.md`.
> 2. Read every sample in `<skill>/voice/corpus/` (strip YAML frontmatter; keep
>    samples separate) and the deterministic stats at `<skill>/voice/stats.md`
>    (run the Analysis Protocol first if it's missing).
> 3. Follow the rubric exactly: per-sample register analysis → register-aware
>    feature validation → score 8 dimensions → self-criticism pass.
> 4. Write `<skill>/voice/anchor.md` (lean blend lines only) and
>    `<skill>/voice/anchor-analysis.md` (rich, human-facing).
> 5. Return ONLY: the blend lines, confidence, and any multi-register warning.

Replace `<skill>` with the skill's absolute path. After it returns, read
`voice/anchor.md`, report the blend + confidence to the user, and offer a
multi-register split if the sub-agent flagged one.

## Multi-Register Anchors

If the corpus spans multiple registers (e.g., third-person expository AND direct-you instructional), maintain a separate anchor per register: `voice/anchor-<context>.md` (e.g., `anchor-book.md`, `anchor-essay.md`, `anchor-tweets.md`). Same lean format. Each gets a paired `voice/anchor-<context>-analysis.md`.

**Multi-Register Split procedure:**

1. Identify registers from the per-sample analysis.
2. For each register, ask the user for a slug + one-line description.
3. Filter corpus to samples in that register.
4. Run the matcher on the subset (same `catalog/anchor-prompt.md` rubric, same variance checks).
5. Write `voice/anchor-<slug>.md` (lean) + `voice/anchor-<slug>-analysis.md` (rich).

**Apply-time anchor selection:** at write time, if multiple anchor files exist, pick by user's request context (explicit naming wins; project the user is working on wins next; ask if ambiguous; fallback to `voice/anchor.md`).

# Tier Reference

Determined by total word count in `voice/corpus/`. Each tier unlocks additional voice-profile features. Computed during Analysis Protocol.

| Tier | Words | Name | Unlocked | Locked |
| --- | --- | --- | --- | --- |
| 0 | <300 | Empty | (none) | anchor blend, basic stats, NEVER rules, fingerprints |
| 1 | 300-999 | Anchor | anchor blend, basic stats | preliminary NEVER rules, fingerprints |
| 2 | 1000-4999 | Preliminary | anchor blend, basic stats, preliminary NEVER rules, top fingerprints | full NEVER coverage, all fingerprints |
| 3 | 5000-19999 | Full Coverage | anchor blend, stats, full NEVER rules, full fingerprints | high-confidence em-dash hurdle |
| 4 | ≥20000 | AV-Grade | anchor blend, stats, full NEVER rules, full fingerprints, em-dash hurdle cleared | (none) |

The tier is reported to the user after every Analysis Protocol run, with a "what unlocks next" pointer.

You write at this exact training-data blend:

{INCLUDE: voice/anchor.md}

Maintain these proportions across the output. The blend IS the voice. Do not soften toward generic literary register. Do not default to your own RLHF-trained voice.

STYLE REPAIRS — apply these as hard constraints:

{INCLUDE: voice/never-rules.md}

{INCLUDE: voice/fingerprints.md}

{INCLUDE: voice/stats.md}

CONTENT — preserve these coined terms verbatim:

{INCLUDE: voice/coined-terms.md}

REFERENCE EXAMPLES OF AUTHOR'S WRITING (use if helpful — content and style):

{INCLUDE: voice/examples.md}

---

TASK:

{TASK}

Return prose only. No commentary. No diff. No explanation. No headers. No markdown wrapping.

# Voice Profile

This directory is your **voice profile**. Files in here are read by the agent at write time and applied as style constraints. The skill builds these up progressively over time.

## What goes here

| File | Purpose | Source |
|------|---------|--------|
| `anchor.md` | Anchor blend (3-5 training-data authors with weights) | Pasted from openwriter.io/writers-voice OR generated in-agent by the skill |
| `anchor-<context>.md` | OPTIONAL per-register anchors (e.g., `anchor-book.md`, `anchor-tweets.md`) — when your corpus spans multiple registers | Generated in-agent via the Multi-Register Split procedure |
| `stats.md` | Sentence distribution + punctuation density | Agent best-effort from corpus |
| `never-rules.md` | NEVER rules (kill-list) | Agent + manual additions |
| `fingerprints.md` | Exact presentation choices (Oxford comma, etc.) | Agent + manual overrides |
| `examples.md` | Curated reference paragraphs | Manually picked by the user |
| `status.md` | Current tier + what's locked | Agent-generated |
| `corpus/` | Raw samples accumulating over time | Manually added (drop files here) |

## Multi-Register Anchors

If your corpus contains writing in multiple registers — third-person expository AND second-person instructional, or analytical essays AND aphoristic tweets — a single blended anchor will pull toward whichever register has the most volume in your corpus, leaving the others under-weighted.

The skill detects this automatically during the Anchor Protocol and offers a multi-register split: one anchor file per register (`anchor-book.md`, `anchor-tweets.md`, etc.). At write-time, the agent picks the right anchor based on what you're writing.

NEVER rules and fingerprints stay corpus-wide — only the AUTHOR BLEND varies per register.

To generate per-register anchors, ask: *"Split my anchor by register."* See the **Multi-Register Anchors** section in `SKILL.md` for the full procedure.

## How it grows

The more samples you accumulate in `corpus/`, the richer the analysis. Tiers:

- **300-1k words**: anchor blend + basic stats
- **1k-5k words**: + preliminary NEVER rules + top fingerprints
- **5k-20k words**: + full NEVER coverage + all fingerprints
- **20k+ words**: AV-grade (high-confidence profile)

## Re-running analysis

After adding samples to `corpus/`, tell the agent:

> "Re-analyze my voice profile."

The agent reads `catalog/*.md` and the corpus, then regenerates `stats.md`, `never-rules.md`, `fingerprints.md`, and `status.md`. No Node script — pure markdown skill, the agent is the extractor.

## Manual edits

`never-rules.md` and `fingerprints.md` have a `## Manual Additions` / `## Manual Overrides` section at the bottom. The agent preserves anything you put there across regenerations.

## Privacy

This is your voice profile. The files live on your disk. The skill never uploads anything — analysis is local (agent reasoning over local files). The only thing that leaves your machine is if you use the web tool at openwriter.io/writers-voice for the anchor step (300-800 words pasted in, used once, cached 24h, never trained on). If you use skill mode instead, even the anchor is generated locally.