# Deep Interview Layer

Apply the deep interview protocol on top of the baseline questions above. Assumption probing and contradiction tracking always run. Research-backed challenges and second-order effects run when the scope warrants it (multi-system changes, infrastructure decisions, technology selection).

**Assumption probing:** After each substantive answer, identify what the user assumed but didn't state. "You described X -- are you assuming Y is already in place?" Surface hidden dependencies and unstated constraints.

**Second-order effects:** For features that touch shared infrastructure or data models, ask what success creates downstream. "If this works and gets adopted, what pressure does it put on [related system]?"

**Research-backed challenges:** Fire background research on technology choices and claims. When findings contradict, challenge directly with citation. When findings support, briefly confirm to build confidence in the decision.

**Contradiction tracking:** If the user's answer contradicts something said earlier, flag it immediately: "Earlier you said X, but this implies Y. Which takes priority?"

**Anti-requirements:** When the user rejects an approach or says "definitely not X," capture the rejection and rationale inline with the related decision. Don't force this -- capture organically when it surfaces.

**Architecture-first ordering:** When several questions remain, ask the ones whose answers would change the *architecture* first -- data model shape, service boundaries, sync vs. async, auth model, storage engine. A question whose answer only affects a label, copy string, or default value can wait or take a reasonable default. Front-loading architecture-changing questions means a redirect lands before the design is built around a wrong assumption, not after.

**Question clustering:** When probing a single dimension (e.g., data model, auth flow), ask 2-3 related questions together using AskUserQuestion's multi-question support. Switch to one-at-a-time when jumping between dimensions.

**Completeness assessment:** Track which dimensions have been explored. Before proposing to move to Phase 2, assess coverage and signal confidence: "We've covered purpose, users, and constraints well. Data flow and failure modes are still thin -- want to explore those, or proceed?"

## Rigor Probes for Ambiguous Gaps

When a user answer leaves a gap on evidence, specificity, counterfactual, or attachment, fire ONE open-ended probe per gap — *not* a multiple-choice menu (exception: territory the user can't evaluate, where menus are the right tool — see Blindspot Pass below). Menus signal which axes the agent thinks matter, biasing the user toward those axes; open-ended forces actual observation:

- **Evidence:** "What's the most concrete thing someone's already done about this — paid for it, built a workaround, quit a tool over it?"
- **Specificity:** "Can you name a team you've actually watched hit this, or are you reasoning?"
- **Counterfactual:** "What do teams do today when this breaks — who reconciles?"
- **Attachment:** "What's the smallest version that would still prove the bet right, and what's excluded?"

Interleave with narrowing moves; do not stack multiple probes in one turn.

## Blindspot Pass for Unfamiliar Territory

When the user signals they *cannot evaluate* a domain — explicit ("I know nothing about auth", "no idea, you pick") or implicit (deferring the same judgment-domain question to the agent 2+ times) — stop extracting guesses and map the decision surface for them instead.

Distinguish **can't-evaluate** (no basis to choose) from **hasn't-decided** (has a basis, just hasn't picked). Only the first triggers this pass; a user who can weigh the options but is undecided gets the normal probes above. This guard keeps the pass from firing on every open question.

Scope the pass to the unfamiliar territory only, not the whole interview. List 3-7 decisions and hazards the user can't see, each with 2-3 concrete options and a recommended default, and present them with `AskUserQuestion` (load its schema via `ToolSearch select:AskUserQuestion` first if unavailable; in Codex use `request_user_input`; where no blocking tool exists, fall back to numbered options in chat) so the user chooses against real alternatives instead of the agent deciding silently. In a non-interactive context, treat the pass as declined and record the recommended defaults as explicit assumptions the user can override later.

## Integration Check Before Phase 1 Exit

Before exiting Phase 1, mentally combine what the user has stated so far. If stated-A + stated-B + agent-default-C produces a downstream effect the user is unlikely to have tracked (e.g., "if mute lives on the rule AND we don't warn on delete, rule-delete silently loses pause state"), fire one open probe per genuine combination. Phase 2.5's call-outs are a safety net for residuals — *not* a punt list for consequences you should have surfaced here.

# Design capture and handoff

Read when writing or reviewing a durable design. The entry point’s authority boundary also applies to commits and handoff.

### Phase 3: Capture the Design

Summarize key decisions in a structured format. For each major component, verify isolation and clarity: it must answer "what does it do, how do you use it, what does it depend on?" and be independently understandable and testable. If working in an existing codebase, note which existing patterns to follow and where targeted improvements fit naturally.

**Design Doc:** Save to `docs/brainstorms/YYYY-MM-DD-<topic>-brainstorm.md`. Required sections: What We're Building, Why This Approach, Key Decisions (with rationale), Open Questions, Next Steps. Collapse the Q&A interview log in a `<details>` block. Include YAML frontmatter with `date` and `topic`. Commit to git -- design decisions are project history.

**Settled vs. directive — don't re-litigate.** A decision the user made with the alternative and its trade-off in view is **settled**: record it in Key Decisions with its rationale and carry it forward — do not re-ask it in Phase 3b, at planning, or during work. A cold **directive** (a choice asserted without anyone weighing it — "build it with X") earns exactly **one** in-pipeline challenge (one pass of the Phase 2 ideation lenses against that specific choice), then it too is recorded and not re-challenged at every downstream stage. A settled label never suppresses defect evidence — a real bug or infeasibility found *inside* a settled approach keeps full severity and is surfaced.

### Phase 3b: Spec Self-Review

Run this checklist before presenting the design doc. Any failure returns to Phase 2 or Phase 3, not Phase 4.

- **Placeholder scan**: no TBD, "figure out later", "appropriate error handling", bracketed gaps, or tasks without concrete criteria.
- **Internal consistency**: names, types, and verbs match across sections (no `createOrder()` in one place and `placeOrder()` in another).
- **Scope containment**: every decision traces back to a stated goal; otherwise cut or surface as explicit scope expansion.
- **Ambiguity sweep**: each Key Decision survives "could a reasonable implementer interpret this two ways?"
- **Assumption validation**: every assumption names its validation method ("we assume X — we'll confirm by Y").
- **Value sourcing**: enumerate every value the work must produce, compute, or display, and confirm the spec names each one's source (an input param, a stored field, a derivation from a named value, or a prior decision). A produced value with no named source is an owed design decision — surface it, don't invent it. Judge by positive enumeration, not introspection: "show the user's local day" that never says where the timezone comes from passes every other check yet hides an undecided source.
- **Non-goals present**: the explicit "Not Doing" list exists and is specific.

Silent pass is valid. Clean draft → move to Phase 4.

### Phase 4: Review and Handoff

Present the design doc to the user for approval. The user explicitly confirming the design is the gate to proceed. When invoked via `/ia-brainstorm`, the command handles spec review dispatch and next-step orchestration.

**Explicit headless mode:** return the design and unresolved assumptions to the caller. Handoff may continue only within the caller's delegated authority; report the design as caller-delegated, not user-approved.

## Anti-Patterns to Avoid

| Anti-Pattern | Better Approach |
|--------------|-----------------|
| Asking 5 questions at once | Ask one at a time across dimensions; cluster 2-3 within a dimension |
| Jumping to implementation details | Stay focused on WHAT, not HOW |
| Proposing overly complex solutions | Start simple, add complexity only if needed |
| Ignoring existing codebase patterns | Research what exists first |
| Making assumptions without validating | State assumptions explicitly and confirm |
| Creating lengthy design documents | Keep it concise--details go in the plan |

## Success Criteria

- Design doc saved to `docs/brainstorms/YYYY-MM-DD-<topic>-brainstorm.md`
- Interactive mode: user approves the spec before handoff. Explicit headless mode: caller-delegated decisions and remaining assumptions are identified.
- All open questions resolved or explicitly deferred with rationale

## Integration

Brainstorming answers WHAT to build. Planning answers HOW. When brainstorm output exists, `/ia-plan` (Claude Code) or the ia-planning skill detects it and skips idea refinement.

- **Next step:** planning, always (`/ia-plan` in Claude Code; the `ia-planning` skill elsewhere)
- **Threat modeling:** when the brainstorm involves auth, payments, external API surfaces, or multi-tenant data, suggest a `ia-security-sentinel` threat model before moving to planning. Catching trust boundary issues at the design stage prevents costly rework.
- **Predecessor:** user request or ambiguous feature description

# Interview and approach selection

Read when requirements need dialogue or multiple approaches need comparison. Headless execution follows the entry point’s caller-delegated decision scope.

### Phase 1: Understand the Idea

**User context calibration (before diving into the idea):**

Read signals from the user's first message to calibrate communication register:
- **Vocabulary**: Are they using technical terms (API, schema, migration) or describing experiences (it's slow, it breaks when...)?
- **Framing**: Are they describing a solution ("build a dashboard") or a problem ("I can't see what's happening")?
- **References**: Are they pointing to code, files, and patterns, or to analogies and comparisons ("something like Notion")?

Adjust question style accordingly. Technical users get architecture-level probing. Non-technical users get experience-level probing. Don't ask about this calibration -- just do it. If signals are ambiguous, default to the vocabulary the user is already using.

**Explore project context first:** Before asking questions, read existing files, docs, and recent commits related to the idea. Understanding what exists prevents asking questions the codebase already answers and grounds the conversation in reality. When the user's wording conflicts with what the code verifiably does ("the retry queue" when nothing retries; a table or endpoint named that doesn't exist), surface the conflict before treating the wording as settled -- silently adopting either side buries a requirements error.

Ask questions **one at a time** by default. When probing a single dimension (e.g., data model, auth flow), clustering 2-3 related questions together is acceptable.

**Facts vs decisions (mid-interview):** before asking, classify each candidate question. A fact (which table holds the field, whether an endpoint exists, what a library supports) is answered by inspecting code or docs, or a quick background lookup, not by asking the user. Reserve the blocking question tool for genuine trade-offs and preferences.

**Premature solutions:** when the user proposes a solution before the requirements are understood, acknowledge it in one line and redirect to the requirement it serves; hold it as a candidate for Phase 2 rather than adopting it. Once Phase 2 has started, evaluate it alongside the other approaches instead of redirecting.

**Info-dump gate (when user offers rich context up-front):** if the user's first message is substantial (>200 words, or dumps requirements in stream-of-consciousness), resist the urge to ask questions one-at-a-time. Instead, respond with 5-10 **numbered clarifying questions** the user can answer in shorthand (`1: yes, 2: channel #ops, 3: no because backwards compat`). Pick questions that remove ambiguity, not questions that show you read the dump. Exit this batched mode when the user's answers show they can be asked about edge cases without basics being explained back to them.

Example after a spec dump:

```
Before I propose approaches, quick clarifications:

1. Auth — SSO (which provider?) or username/password?
2. Sync or async for the webhook delivery?
3. Which of the three integrations is P0?
4. "Fast enough" in the spec — what's the actual number?

Answer whichever you know; leave blanks for the rest.
```

**Question Techniques:**

1. **Prefer multiple choice when natural options exist.** Good: "Notification: (a) email, (b) in-app, (c) both?" Avoid: "How should users be notified?"
2. **Start broad, then narrow.** Core purpose → users → constraints.
3. **Validate assumptions and probe success early.** "I'm assuming users are logged in — correct?" / "How will you know this is working?"

**Key Topics to Explore:**

| Topic | Example Questions |
|-------|-------------------|
| Purpose | What problem does this solve? What's the motivation? |
| Users | Who uses this? What's their context? |
| Constraints | Any technical limitations? Timeline? Dependencies? |
| Success | How will you measure success? What's the happy path? |
| Edge Cases | What shouldn't happen? Any error states to consider? |
| Existing Patterns | Are there similar features in the codebase to follow? |
| Non-goals | What is explicitly NOT in scope? |

See [deep-interview.md](./deep-interview.md) for deep interview techniques, including **rigor probes** (evidence/specificity/counterfactual/attachment as open-ended forced production, not menus), the **blindspot pass** for domains the user can't evaluate, and the **integration check** that fires before Phase 1 exit when combining stated answers + agent defaults produces an unsurfaced downstream effect.

**Exit Condition:** Continue until the idea is clear OR user says "proceed". Before moving to Phase 2, summarize understanding in 3-5 bullets and confirm with the user.

### Phase 2: Explore Approaches

After understanding the idea, propose 2-3 concrete approaches.

**Structure for Each Approach:**

```markdown
### Approach A: [Name]

[2-3 sentence description]

**Pros:**
- [Benefit 1]
- [Benefit 2]

**Cons:**
- [Drawback 1]
- [Drawback 2]

**Best when:** [Circumstances where this approach shines]
```

**Guidelines:**
- Lead with a recommendation and explain why
- Be honest about trade-offs
- Consider YAGNI--simpler is usually better
- Reference codebase patterns when relevant
- If no approach is accepted after 2 rounds, ask the user to describe their preferred direction directly

**Ideation lenses** (use 2-3 to stress-test approaches when the design space is wide):
- **Inversion**: What if we solved the opposite problem?
- **Constraint removal**: What would we build if [biggest constraint] didn't exist?
- **Simplification**: What's the version that ships in a day?
- **10x version**: What if this needed to handle 10x the scale?
- **Expert lens**: How would [domain expert] approach this?

**"Not Doing" list:** Include an explicit list of what the chosen approach will NOT do. Focus is about saying no to good ideas. Make the trade-offs visible so they're a deliberate choice, not an oversight.

**Assumptions with validation:** For each key assumption in the chosen approach, state how to test it. Not just "we assume X" but "we assume X -- we'll know by [validation method]."

# Pre-write scope synthesis

Read after substantive dialogue or when documenting a Standard/Deep design. Interaction mode and authority come from the skill entry point.

### Phase 2.5: Pre-Write Scope Synthesis

Surface the scope interpretation so the user can correct it before Phase 3 writes the design doc. Phase 2.5 catches scope misalignment before the doc is written; Phase 3b catches drafting issues after.

**Two-stage shape: internal draft, then chat-time scoping synthesis.** Compose in two stages. Stage 1 is an internal three-bucket thinking pass (Stated / Inferred / Out of scope) for comprehensive scope analysis. Stage 2 is what the user sees — shaped like what two product collaborators would confirm before writing a PRD. The internal draft never reaches the user verbatim; it routes into the Phase 3 doc body.

**Stage 1 — internal three-bucket draft (thinking, not output):**
- **Stated** — what the user said directly. Explicit user-language anchors.
- **Inferred** — gaps the agent filled with assumptions. Most actionable bucket; bets the user can correct.
- **Out of scope** — deliberately excluded items.

Use this as a thinking step. Do not paste it into chat.

**Stage 2 — user-facing scoping synthesis.** Up to four named sections, each render-conditional. Empty sections are omitted, not padded:

1. **What we're building** (always present) — 1-3 sentences. The shape that emerged from dialogue, forward-looking, plain words. Not a transcript of "you said X".
2. **Key trade-offs** (conditional) — 1-3 bullets, each with a brief why. Render only when real trade-offs were made.
3. **What's not in scope** (conditional) — 1-3 bullets, or fold into a sentence. Render only when deferred items would surprise a downstream reader if absent.
4. **Call-outs** (conditional) — 0-3 bullets. Residual forks the dialogue didn't resolve: post-dialogue consequences, silent agent inferences, or — in pre-loaded contexts — scope bets the user is seeing for the first time. Not "questions the agent could have asked during Phase 1 but didn't" — if a call-out reads like a missed dialogue question, Phase 1's integration check failed; flag the gap.

Close with: *"Confirm and I'll write the design doc next. Or tell me what to change."*

**Path A vs Path B gate.** Routing depends on TWO signals: (1) did any *blocking* question fire before Phase 2.5? AND (2) what tier did Phase 0 classify? Blocking questions = scope disambiguation, dialogue probes, approach selection menus. Internal classification and pressure-tests do not count.

- **Path A** — Lightweight tier AND no blocking questions fired → announce-mode. Emit "What we're building" prose only (no other sections, no confirmation question), then proceed to Phase 3 doc-write in the same turn. Lightweight Path A docs are short; post-hoc revision is cheap.
- **Path B** — Standard/Deep tier OR any blocking question fired → full synthesis with confirmation gate. Two scenarios fire Path B: the user invested answer-time in dialogue, or pre-loaded substantive scope content. Either way, the substance earns a real checkpoint. The tier guard catches pre-loaded Deep brainstorms that would otherwise shortcut via the no-questions branch.

**Keep tests per section.** Each conditional section has its own keep test; failing items dissolve into the internal draft only.
- **Trade-offs**: would the user be surprised if I didn't surface this acknowledgment? Mechanical or inevitable choices fail.
- **Deferred**: is a reasonable downstream reader likely to ask "why isn't X here?" Mechanical excludes fail.
- **Call-outs**: two-step test. (1) Affirmability: would the user need to read code to evaluate this? If yes, it's doc-body content — cut. (2) Keep only if it's a real scope fork, non-obvious inclusion/exclusion, cheap-now-expensive-later correction, or non-obvious consequence of combined multi-turn answers. (3) Phase 1 boundary: if the call-out depends only on Phase 1 facts (no Phase 2 approach, no later-surfaced default), Phase 1's integration check failed — cut and revisit Phase 1. Call-outs catch what Phase 1 *couldn't* surface, not what it *should have*.

Cut re-statements of Q&A turns, re-statements of the picked Phase 2 approach, mechanical items, and implementation choices that settle during planning.

**Bullet budget across sections 2-4 combined.** Heuristic, not law — the real discipline is each section's keep test:

| Tier | Typical total | Hard ceiling |
|---|---|---|
| Lightweight | 0-1 | 2 |
| Standard | 2-4 | 5 |
| Deep | 3-7 | 9 |

Above the ceiling means the synthesis is mis-shapen — re-cut at a higher level of abstraction, do not raise the cap.

**Detail level: conversational, not documentary.** 1 line ideally, 2 max. Bullets that need semicolons stringing clauses or an internal list are two decisions sharing a bullet — split or drop.

**Re-present after revision; write only on confirm.** If the user revises any bullet (even trivially), integrate the change, re-present, and wait for explicit confirmation. A revision is not a confirmation.

**Explicit headless mode:** compose the synthesis without requesting confirmation. Route inferred items to `## Assumptions` in the Phase 3 doc with validation methods; do not label them user-approved decisions. Stated requirements and non-goals retain their source. Report any decision beyond the caller's delegated authority instead of silently choosing it.

Skip Phase 2.5 entirely when Phase 0.2 detected requirements were already clear and the flow proceeded straight to summary without a Phase 1 dialogue. Path A handles every other Lightweight case.