proof-points
Builds and maintains a sourced, gated registry of the proof points your messaging relies on (metrics, customer quotes, case study results, analyst citations), then deepens brain Section 6. Use when pulling proof points out of a deck or battlecard, checking whether a stat is safe
Install
npx skills add https://github.com/stefanoskarakasis/Product-Marketing-Skills/tree/main/pmm-positioning/skills/proof-points
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install stefanoskarakasis-product-marketing-skills@llmmart
git clone https://github.com/stefanoskarakasis/Product-Marketing-Skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole stefanoskarakasis/product-marketing-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Proof Points
How This Works
Every stat, quote, and case study in your messaging is a promise someone can ask you to back up. Most companies have a mash of decks, battlecards and case study PDFs repeating stale numbers. This skill is both the intake and the gate: it mines claim
candidates out of that mess, then makes sure nothing gets called "approved"
without a source, no customer name ships without documented permission to use
it, and every forbidden claim carries a stated reason — not a bare
prohibition nobody can act on. It deepens your brain's existing Section 6 in
place, the same Proof Points Registry message-house and
positioning-messaging already read.
Trigger
- When: Building a proof points registry from scratch out of existing sales decks, battlecards, case studies, or pricing pages; adding a single new metric, quote, or case study directly; checking whether a specific stat or phrase is safe to use before it ships in copy; or auditing brain Section 6 for unsourced claims, stale proof, undocumented customer names, or unsupported superlatives.
- Not for: Structuring raw, unstructured customer conversation — a call
recording or interview transcript nobody's synthesized yet — into
discovery insights →
interview-summary, run first. That's a qualitatively different task (JTBD synthesis of what customers said) from this skill's Extract mode, which mines already-written claims out of already-somewhat-structured GTM collateral (a deck, a battlecard, a case study doc) — if Extract mode turns up a customer quote that traces back to a raw call nobody's written up, route that call throughinterview-summaryfirst, then bring the resulting quote back here. Vocabulary, tone, and forbidden language (not claims) →brand-voice. Building the positioning statement or messaging hierarchy itself →positioning-messaging, run this first so it has a gated registry to cite from. Packaging Section 6 evidence into a rep-ready handoff document →product-messaging-playbook, run this first so it pulls clean, sourced proof. - Example prompts:
- "Pull proof points out of this battlecard"
- "We don't have a registry — start one from these 3 sales decks"
- "Add a proof point: 40% faster onboarding, from our Q3 customer survey"
- "Can I say 'industry-leading' in this deck?"
- "Audit our proof points — anything stale or unsourced?"
Inputs
- Args: For Extract mode — one or more source documents or pasted text (a sales deck, battlecard, case study, pricing page, one-pager). For Add mode — a single claim with its source. For Verify mode — a claim to check. For Audit mode — no additional args beyond "audit." If the mode isn't clear from what's given, ask.
- Defaults: If brain Section 6 already has entries but they're missing the layers this skill requires (source, approval-to-use, staleness date), treat this as additive — fill the missing layers on the existing entries rather than re-litigating claims that are already sourced and current. If Section 6 is empty or near-empty and the user hasn't specified a mode, ask whether they have source material to extract from before assuming claims will be added one at a time — most first-time registries start from Extract mode, not Add mode.
- Context keys:
/foundation/brain.md— optional (created on the first confirmed write if missing). Section 6 (Proof Points Registry). Section 1 (Product Context) read for plausibility only — never to invent a claim the user didn't supply.- Brain contract: Reads Section 6, Section 8 (
prohibited,approvals), and Section 1 for context only. Writes Section 6 only — approved metrics, customer quotes/case studies, and forbidden claims, appended on explicit confirmation. Never writes to any other section. If no brain exists, creates/foundation/brain.mdholding Section 6 alone (cold start).
Connectors (Optional)
| Connector | What it adds |
|---|---|
~~call recordings |
Customer quotes (enter as unverified) |
~~CRM |
Metric candidates (enter as unverified) |
~~product analytics |
Usage evidence (paste export) |
~~team chat |
Win announcements and customer praise (enter as unverified) |
~~cloud storage |
Case studies and decks in Drive |
No connectors? Paste the material and I'll work from that. The output has the same shape. Pulled facts are drafts: each is tagged with category, tool and date, and nothing is written until you confirm.
Pre-flight
- Load
/foundation/brain.mdSection 6 if it exists. Also read Section 8prohibitedandapprovalsif present: a claim that touches a prohibited topic is flagged and not added as approved, and each customer number or quote notes the approver Section 8 names. - Brain absent: continue, don't block. This skill works from the evidence you give it. Say once: "No PMM brain found, so I'll work from your material and start the brain with this section only. Run
product-marketing-contextlater (a separate plugin) to build the rest." The write step handles the cold start. - Soft block: request doesn't clearly state Extract, Add, Verify, or Audit → ask which applies. If Section 6 is empty and the user just says "build our proof points," default the question toward Extract mode first: "Do you have decks, battlecards, or case studies I can pull from, or are you adding claims one at a time?"
- Soft block (Add mode only): a claim offered with no source → do not
reject it outright; proceed to Step 3, where it gets flagged
[NEEDS PROOF]instead of silently marked approved.
Steps
Step 0 — Load Context
Load brain Section 6 as it currently exists, and Section 1 for product-name context only.
Gate check: brain absent → surface: "Brain not found. Run
product-marketing-context first — even a thin Section 6 gives this skill
something real to work from."
Step 1 — Classify the Request
Determine which of four modes applies. If ambiguous, ask:
"Do you want to pull proof points from existing material (decks, battlecards, case studies), add one claim directly, check whether a specific claim is safe to use, or audit the existing registry?"
Step 2 — Extract Mode: Mine Claims from Source Material
This is the on-ramp for the common case: no clean registry exists, only a mash of sales decks, battlecards, one-pagers, and case study documents.
- Read through every piece of source material provided and pull out every claim-shaped statement: a number or stat, a named customer outcome, a direct quote, a case study reference, or a superlative/comparison claim ("industry-leading," "fastest," "the only platform that...").
- Classify each extracted candidate by type — Metric, Customer quote, Case study result, Analyst citation, or Superlative — same taxonomy Add mode uses.
- Every extracted claim starts
[NEEDS PROOF], regardless of whether the source material itself cites a source. A deck slide that says "per Q3 earnings" is not the same as a confirmed-current source — decks get copied forward for quarters after the number goes stale. If the source material does cite something, carry that citation through as a proposed, unconfirmed source attached to the flag:[NEEDS PROOF — deck cites Q3 2025 earnings, unconfirmed]. This gives the user something concrete to confirm or reject rather than a blank slate, without treating an in-deck citation as verified provenance on its own. - A customer name or quote pulled from a deck gets
[NEEDS APPROVAL]in addition to[NEEDS PROOF]— sales enablement material routinely reuses customer names before (or after the expiry of) actual legal/customer sign-off. A name appearing in a deck is not documented permission to use it. - When the same claim (or a near-duplicate phrasing) appears across multiple source documents, surface it once as a single candidate, noting how many source documents it appeared in — never register the same claim twice because it showed up in two decks.
- Return the full candidate list before writing anything, grouped by type:
"Found [N] claim-shaped statements across [source documents]. All flagged
[NEEDS PROOF]until confirmed current — here they are: [list]. Register all of them as flagged candidates, or go through and confirm/reject first?" - On confirmation, batch-append to Section 6 via Step 6 — still flagged. Extraction never promotes a claim to approved on its own; that only happens later, through Add mode's explicit source confirmation or Verify mode, once the user has actually checked the source is current.
Step 3 — Add Mode: Register a New Claim Directly
Classify the claim by type — Metric, Customer quote, Case study result, or Analyst citation — and require a source for each:
| Claim type | What "sourced" means |
|---|---|
| Metric | Where the number comes from and when it was pulled (internal dashboard, public filing, survey — with date) |
| Customer quote | Who said it, their role and company, and confirmation they've approved being named — not just that they said it on a call |
| Case study result | A published or internally-approved case study the number traces back to |
| Analyst citation | The named report, firm, and publication date |
No source → the claim is registered as [NEEDS PROOF], not silently
promoted to approved. A customer quote with no documented approval-to-use
gets [NEEDS APPROVAL] even if the quote itself is real — a customer
saying something on a call is not the same as a customer agreeing to be
named in your marketing.
Reject vague quantifiers as claims on their own: "many customers report..." is not a proof point. Either name the number and its source, or don't register it yet.
Step 4 — Verify Mode: Check a Claim Before It Ships
User supplies a specific stat or phrase they want to use. Check it against the current registry and return one of three answers:
- Approved — it's in the registry with a source (and approval-to-use, if it names a customer). Cite the source back to the user.
- Forbidden — it matches or resembles an entry in the Forbidden Claims list. State which entry and why (see Step 5's Forbidden Claims layer, built or updated in either Extract or Add mode).
- Not yet registered — it isn't in the registry either way. This is not automatic approval or automatic rejection — route to Extract mode (if it likely lives in existing collateral) or Add mode (if the user has the source in hand) rather than flagging it unusable with no path forward.
Superlatives ("industry-leading," "best-in-class," "the only platform that...") without a named comparison are flagged as unsupported even if no Forbidden Claims entry names them specifically — a superlative with no comparison is a claim with no evidence, whether or not it's been banned outright.
Step 5 — Audit Mode: Review the Existing Registry
Walk every entry currently in Section 6 — not a sample — and flag:
- No source — mark
[NEEDS PROOF]. - Customer name with no documented approval-to-use — mark
[NEEDS APPROVAL]. - Stale — a dated metric or citation older than 12 months with no
refresh noted. Flag
[STALE — recheck], don't silently drop it. - Unsupported superlative — language claiming "best," "most," "only," or similar with no named comparison anywhere in the entry.
Also build or update the Forbidden Claims list here: every forbidden claim requires a stated reason — legal exposure, unproven, competitively risky, or previously retracted — never a bare "don't say this."
**Forbidden Claims:**
- "[claim or claim pattern]" — [reason: legal / unproven / competitively
risky / retracted]. [One sentence of context if useful.]
Return the full flagged list as the punch list — this is the output of Audit mode; it does not also rebuild the registry from scratch.
Step 6 — Deepen Brain Section 6 (on Confirmation)
Show the exact addition before writing:
"Adding to brain Section 6 — [new approved metric(s) / flagged candidate(s) / customer quote(s) / case study result(s) / forbidden claim(s)], sourced from [source]. Existing content stays unchanged. Here's what I'll append: [show exact text]. Save this?"
Append to /foundation/brain.md Section 6 — do not replace existing
content, do not create a separate file:
## Section 6: Proof Points Registry
[existing content unchanged]
### Registry Update (added via proof-points, [date])
**Approved Metrics:**
- [claim] — [source, date] `[NEEDS PROOF]` if unsourced
**Customer Quotes & Case Study Results:**
- "[quote]" — [name, role, company] ([approval-to-use confirmed / `[NEEDS APPROVAL]`])
**Forbidden Claims:**
- "[claim]" — [reason]
**Extracted this session (if Extract mode ran):**
- [candidate] — [claim type] `[NEEDS PROOF — <proposed source if the
material cited one, else blank>]`, found in [source document(s)]
**Audit flags (if this session ran Audit mode):**
- [entry] — [NEEDS PROOF / NEEDS APPROVAL / STALE / unsupported superlative]
Cold start (no brain file yet). If /foundation/brain.md does not exist, create it after the same confirmation, containing only this section:
# PMM Brain — [product name, or "not set"]
🟡 Cold start — built via proof-points on [date]. Run `product-marketing-context` to build the other sections.
## Section 6: Proof Points Registry
[the addition shown above]
Do not invent content for the other sections. If this environment cannot write files, show the block and tell the user to save it as brain.md in their working folder or Project knowledge so later sessions can read it. Never say it was saved unless the write succeeded.
Never write without this explicit confirmation.
Step 7 — Learning Close
Append one entry to /context/skill-sessions.md, in the canonical format
defined in product-marketing-context/.claude-plugin/skill-sessions-format.md
(Type A — execution session):
type: execution
skill: proof-points
session_date: [YYYY-MM-DD]
pattern: [one falsifiable statement about this session, or "none"]
source: [surprised / wrong / missing / N/A]
Write directly, no permission needed — separate from the brain write above, which still requires explicit confirmation.
Outputs
- Files written:
/foundation/brain.mdSection 6 — appended approved metrics, extracted candidates, customer quotes/case studies, and forbidden claims (Step 6), only after explicit confirmation./context/skill-sessions.md— one appended entry per session (Step 7). - Chat output format: Mode stated (Extract / Add / Verify / Audit) → candidate list, findings, or new registry entries → confirmation prompt for the brain write.
- External side effects: N/A
- Next skill: After brain Section 6 is deepened, check
product-marketing-context/.claude-plugin/next-skill-map.mdfor "After proof-points" and surface that prompt. Do not auto-run — ask.
Operating Rules
The source, approval and confirmation rules live in the Steps and are tested in the Quality Gate. These are stated only here.
- No source means not approved, full stop. A plausible number with no traceable origin is a liability.
- Extraction never grants approval. A claim pulled from a deck is a candidate even when the deck cites a source; that citation is carried as unconfirmed.
- Customer names need documented approval-to-use. A customer saying something to a rep, or appearing in a battlecard, is not agreement to be quoted publicly.
- Never silently promote an unsourced or extracted claim. Flag it and move on.
- Deepen brain Section 6, never fork a second claims file.
message-houseandpositioning-messagingexpect it to be the single source of truth for proof. - Superlatives get flagged, not auto-blocked. The decision stays with the user.
Quality Gate
| Check | Standard | Pass = |
|---|---|---|
| Every approved claim sourced | No claim marked approved without a cited source | Yes |
| Extracted claims start flagged | Every Extract-mode candidate registered [NEEDS PROOF] regardless of in-deck citation |
Yes |
| Customer approval-to-use documented | Every named customer quote/case study has confirmed permission, not just evidence they said it or appeared in a deck | Yes |
| Forbidden claims carry a reason | Every entry states legal / unproven / competitively risky / retracted | Yes |
| Duplicates merged | The same claim across multiple source documents is registered once, with occurrence count | Yes |
| Audit covers full registry | Every existing Section 6 entry reviewed, not a sample | Yes |
| Superlatives flagged | Unsupported superlatives caught even with no matching Forbidden Claims entry | Yes |
| Existing content preserved | Prior Section 6 content untouched unless explicitly revised | Yes |
| Brain write confirmed | Exact addition (including bulk Extract-mode writes) shown and confirmed before writing | Yes |
| Cold start handled | With no brain file the skill proceeded without redirecting; the only file created held Section 6 alone (or the block was shown for pasting) | Yes |
| Learning Close ran | /context/skill-sessions.md has a new type: execution row for this session |
Yes |
Files (product-marketing-skills)
-
evals
-
proof-points.eval.md 9.1 KB
Loading…
-
-
SKILL.md 17.6 KB
--- name: proof-points description: > Builds and maintains a sourced, gated registry of the proof points your messaging relies on (metrics, customer quotes, case study results, analyst citations), then deepens brain Section 6. Use when pulling proof points out of a deck or battlecard, checking whether a stat is safe to use, sourcing a new claim before it ships, or auditing claims for anything unsourced, stale or unapproved. --- # Proof Points ## How This Works Every stat, quote, and case study in your messaging is a promise someone can ask you to back up. Most companies have a mash of decks, battlecards and case study PDFs repeating stale numbers. This skill is both the intake and the gate: it mines claim candidates out of that mess, then makes sure nothing gets called "approved" without a source, no customer name ships without documented permission to use it, and every forbidden claim carries a stated reason — not a bare prohibition nobody can act on. It deepens your brain's existing Section 6 in place, the same Proof Points Registry `message-house` and `positioning-messaging` already read. --- ## Trigger - **When:** Building a proof points registry from scratch out of existing sales decks, battlecards, case studies, or pricing pages; adding a single new metric, quote, or case study directly; checking whether a specific stat or phrase is safe to use before it ships in copy; or auditing brain Section 6 for unsourced claims, stale proof, undocumented customer names, or unsupported superlatives. - **Not for:** Structuring raw, unstructured customer conversation — a call recording or interview transcript nobody's synthesized yet — into discovery insights → `interview-summary`, run first. That's a qualitatively different task (JTBD synthesis of what customers said) from this skill's Extract mode, which mines already-written claims out of already-somewhat-structured GTM collateral (a deck, a battlecard, a case study doc) — if Extract mode turns up a customer quote that traces back to a raw call nobody's written up, route that call through `interview-summary` first, then bring the resulting quote back here. Vocabulary, tone, and forbidden *language* (not claims) → `brand-voice`. Building the positioning statement or messaging hierarchy itself → `positioning-messaging`, run this first so it has a gated registry to cite from. Packaging Section 6 evidence into a rep-ready handoff document → `product-messaging-playbook`, run this first so it pulls clean, sourced proof. - **Example prompts:** - "Pull proof points out of this battlecard" - "We don't have a registry — start one from these 3 sales decks" - "Add a proof point: 40% faster onboarding, from our Q3 customer survey" - "Can I say 'industry-leading' in this deck?" - "Audit our proof points — anything stale or unsourced?" --- ## Inputs - **Args:** For Extract mode — one or more source documents or pasted text (a sales deck, battlecard, case study, pricing page, one-pager). For Add mode — a single claim with its source. For Verify mode — a claim to check. For Audit mode — no additional args beyond "audit." If the mode isn't clear from what's given, ask. - **Defaults:** If brain Section 6 already has entries but they're missing the layers this skill requires (source, approval-to-use, staleness date), treat this as additive — fill the missing layers on the existing entries rather than re-litigating claims that are already sourced and current. If Section 6 is empty or near-empty and the user hasn't specified a mode, ask whether they have source material to extract from before assuming claims will be added one at a time — most first-time registries start from Extract mode, not Add mode. - **Context keys:** - `/foundation/brain.md` — optional (created on the first confirmed write if missing). Section 6 (Proof Points Registry). Section 1 (Product Context) read for plausibility only — never to invent a claim the user didn't supply. - **Brain contract:** Reads Section 6, Section 8 (`prohibited`, `approvals`), and Section 1 for context only. Writes Section 6 only — approved metrics, customer quotes/case studies, and forbidden claims, appended on explicit confirmation. Never writes to any other section. If no brain exists, creates `/foundation/brain.md` holding Section 6 alone (cold start). --- ## Connectors (Optional) | Connector | What it adds | |---|---| | `~~call recordings` | Customer quotes (enter as unverified) | | `~~CRM` | Metric candidates (enter as unverified) | | `~~product analytics` | Usage evidence (paste export) | | `~~team chat` | Win announcements and customer praise (enter as unverified) | | `~~cloud storage` | Case studies and decks in Drive | > **No connectors?** Paste the material and I'll work from that. The output has the same shape. > Pulled facts are drafts: each is tagged with category, tool and date, and nothing is written until you confirm. --- ## Pre-flight - Load `/foundation/brain.md` Section 6 if it exists. Also read Section 8 `prohibited` and `approvals` if present: a claim that touches a prohibited topic is flagged and not added as approved, and each customer number or quote notes the approver Section 8 names. - **Brain absent: continue, don't block.** This skill works from the evidence you give it. Say once: "No PMM brain found, so I'll work from your material and start the brain with this section only. Run `product-marketing-context` later (a separate plugin) to build the rest." The write step handles the cold start. - **Soft block:** request doesn't clearly state Extract, Add, Verify, or Audit → ask which applies. If Section 6 is empty and the user just says "build our proof points," default the question toward Extract mode first: *"Do you have decks, battlecards, or case studies I can pull from, or are you adding claims one at a time?"* - **Soft block (Add mode only):** a claim offered with no source → do not reject it outright; proceed to Step 3, where it gets flagged `[NEEDS PROOF]` instead of silently marked approved. --- ## Steps ### Step 0 — Load Context Load brain Section 6 as it currently exists, and Section 1 for product-name context only. **Gate check:** brain absent → surface: *"Brain not found. Run `product-marketing-context` first — even a thin Section 6 gives this skill something real to work from."* ### Step 1 — Classify the Request Determine which of four modes applies. If ambiguous, ask: > "Do you want to pull proof points from existing material (decks, > battlecards, case studies), add one claim directly, check whether a > specific claim is safe to use, or audit the existing registry?" ### Step 2 — Extract Mode: Mine Claims from Source Material This is the on-ramp for the common case: no clean registry exists, only a mash of sales decks, battlecards, one-pagers, and case study documents. 1. Read through every piece of source material provided and pull out every claim-shaped statement: a number or stat, a named customer outcome, a direct quote, a case study reference, or a superlative/comparison claim ("industry-leading," "fastest," "the only platform that..."). 2. Classify each extracted candidate by type — Metric, Customer quote, Case study result, Analyst citation, or Superlative — same taxonomy Add mode uses. 3. **Every extracted claim starts `[NEEDS PROOF]`, regardless of whether the source material itself cites a source.** A deck slide that says "per Q3 earnings" is not the same as a confirmed-current source — decks get copied forward for quarters after the number goes stale. If the source material does cite something, carry that citation through as a proposed, unconfirmed source attached to the flag: `[NEEDS PROOF — deck cites Q3 2025 earnings, unconfirmed]`. This gives the user something concrete to confirm or reject rather than a blank slate, without treating an in-deck citation as verified provenance on its own. 4. A customer name or quote pulled from a deck gets `[NEEDS APPROVAL]` in addition to `[NEEDS PROOF]` — sales enablement material routinely reuses customer names before (or after the expiry of) actual legal/customer sign-off. A name appearing in a deck is not documented permission to use it. 5. When the same claim (or a near-duplicate phrasing) appears across multiple source documents, surface it once as a single candidate, noting how many source documents it appeared in — never register the same claim twice because it showed up in two decks. 6. Return the full candidate list before writing anything, grouped by type: > "Found [N] claim-shaped statements across [source documents]. All > flagged `[NEEDS PROOF]` until confirmed current — here they are: > [list]. Register all of them as flagged candidates, or go through and > confirm/reject first?" 7. On confirmation, batch-append to Section 6 via Step 6 — still flagged. Extraction never promotes a claim to approved on its own; that only happens later, through Add mode's explicit source confirmation or Verify mode, once the user has actually checked the source is current. ### Step 3 — Add Mode: Register a New Claim Directly Classify the claim by type — Metric, Customer quote, Case study result, or Analyst citation — and require a source for each: | Claim type | What "sourced" means | |---|---| | Metric | Where the number comes from and when it was pulled (internal dashboard, public filing, survey — with date) | | Customer quote | Who said it, their role and company, and confirmation they've approved being named — not just that they said it on a call | | Case study result | A published or internally-approved case study the number traces back to | | Analyst citation | The named report, firm, and publication date | No source → the claim is registered as `[NEEDS PROOF]`, not silently promoted to approved. A customer quote with no documented approval-to-use gets `[NEEDS APPROVAL]` even if the quote itself is real — a customer saying something on a call is not the same as a customer agreeing to be named in your marketing. Reject vague quantifiers as claims on their own: "many customers report..." is not a proof point. Either name the number and its source, or don't register it yet. ### Step 4 — Verify Mode: Check a Claim Before It Ships User supplies a specific stat or phrase they want to use. Check it against the current registry and return one of three answers: 1. **Approved** — it's in the registry with a source (and approval-to-use, if it names a customer). Cite the source back to the user. 2. **Forbidden** — it matches or resembles an entry in the Forbidden Claims list. State which entry and why (see Step 5's Forbidden Claims layer, built or updated in either Extract or Add mode). 3. **Not yet registered** — it isn't in the registry either way. This is not automatic approval or automatic rejection — route to Extract mode (if it likely lives in existing collateral) or Add mode (if the user has the source in hand) rather than flagging it unusable with no path forward. Superlatives ("industry-leading," "best-in-class," "the only platform that...") without a named comparison are flagged as unsupported even if no Forbidden Claims entry names them specifically — a superlative with no comparison is a claim with no evidence, whether or not it's been banned outright. ### Step 5 — Audit Mode: Review the Existing Registry Walk every entry currently in Section 6 — not a sample — and flag: - **No source** — mark `[NEEDS PROOF]`. - **Customer name with no documented approval-to-use** — mark `[NEEDS APPROVAL]`. - **Stale** — a dated metric or citation older than 12 months with no refresh noted. Flag `[STALE — recheck]`, don't silently drop it. - **Unsupported superlative** — language claiming "best," "most," "only," or similar with no named comparison anywhere in the entry. Also build or update the Forbidden Claims list here: every forbidden claim requires a stated reason — legal exposure, unproven, competitively risky, or previously retracted — never a bare "don't say this." ```markdown **Forbidden Claims:** - "[claim or claim pattern]" — [reason: legal / unproven / competitively risky / retracted]. [One sentence of context if useful.] ``` Return the full flagged list as the punch list — this is the output of Audit mode; it does not also rebuild the registry from scratch. ### Step 6 — Deepen Brain Section 6 (on Confirmation) Show the exact addition before writing: > "Adding to brain Section 6 — [new approved metric(s) / flagged > candidate(s) / customer quote(s) / case study result(s) / forbidden > claim(s)], sourced from [source]. Existing content stays unchanged. > Here's what I'll append: [show exact text]. Save this?" Append to `/foundation/brain.md` Section 6 — do not replace existing content, do not create a separate file: ```markdown ## Section 6: Proof Points Registry [existing content unchanged] ### Registry Update (added via proof-points, [date]) **Approved Metrics:** - [claim] — [source, date] `[NEEDS PROOF]` if unsourced **Customer Quotes & Case Study Results:** - "[quote]" — [name, role, company] ([approval-to-use confirmed / `[NEEDS APPROVAL]`]) **Forbidden Claims:** - "[claim]" — [reason] **Extracted this session (if Extract mode ran):** - [candidate] — [claim type] `[NEEDS PROOF — <proposed source if the material cited one, else blank>]`, found in [source document(s)] **Audit flags (if this session ran Audit mode):** - [entry] — [NEEDS PROOF / NEEDS APPROVAL / STALE / unsupported superlative] ``` **Cold start (no brain file yet).** If `/foundation/brain.md` does not exist, create it after the same confirmation, containing only this section: ```markdown # PMM Brain — [product name, or "not set"] 🟡 Cold start — built via proof-points on [date]. Run `product-marketing-context` to build the other sections. ## Section 6: Proof Points Registry [the addition shown above] ``` Do not invent content for the other sections. If this environment cannot write files, show the block and tell the user to save it as `brain.md` in their working folder or Project knowledge so later sessions can read it. Never say it was saved unless the write succeeded. Never write without this explicit confirmation. ### Step 7 — Learning Close Append one entry to `/context/skill-sessions.md`, in the canonical format defined in `product-marketing-context/.claude-plugin/skill-sessions-format.md` (Type A — execution session): ```yaml type: execution skill: proof-points session_date: [YYYY-MM-DD] pattern: [one falsifiable statement about this session, or "none"] source: [surprised / wrong / missing / N/A] ``` Write directly, no permission needed — separate from the brain write above, which still requires explicit confirmation. --- ## Outputs - **Files written:** `/foundation/brain.md` Section 6 — appended approved metrics, extracted candidates, customer quotes/case studies, and forbidden claims (Step 6), only after explicit confirmation. `/context/skill-sessions.md` — one appended entry per session (Step 7). - **Chat output format:** Mode stated (Extract / Add / Verify / Audit) → candidate list, findings, or new registry entries → confirmation prompt for the brain write. - **External side effects:** N/A - **Next skill:** After brain Section 6 is deepened, check `product-marketing-context/.claude-plugin/next-skill-map.md` for "After proof-points" and surface that prompt. Do not auto-run — ask. --- ## Operating Rules The source, approval and confirmation rules live in the Steps and are tested in the Quality Gate. These are stated only here. - **No source means not approved, full stop.** A plausible number with no traceable origin is a liability. - **Extraction never grants approval.** A claim pulled from a deck is a candidate even when the deck cites a source; that citation is carried as unconfirmed. - **Customer names need documented approval-to-use.** A customer saying something to a rep, or appearing in a battlecard, is not agreement to be quoted publicly. - **Never silently promote an unsourced or extracted claim.** Flag it and move on. - **Deepen brain Section 6, never fork a second claims file.** `message-house` and `positioning-messaging` expect it to be the single source of truth for proof. - **Superlatives get flagged, not auto-blocked.** The decision stays with the user. ## Quality Gate | Check | Standard | Pass = | |---|---|---| | Every approved claim sourced | No claim marked approved without a cited source | Yes | | Extracted claims start flagged | Every Extract-mode candidate registered `[NEEDS PROOF]` regardless of in-deck citation | Yes | | Customer approval-to-use documented | Every named customer quote/case study has confirmed permission, not just evidence they said it or appeared in a deck | Yes | | Forbidden claims carry a reason | Every entry states legal / unproven / competitively risky / retracted | Yes | | Duplicates merged | The same claim across multiple source documents is registered once, with occurrence count | Yes | | Audit covers full registry | Every existing Section 6 entry reviewed, not a sample | Yes | | Superlatives flagged | Unsupported superlatives caught even with no matching Forbidden Claims entry | Yes | | Existing content preserved | Prior Section 6 content untouched unless explicitly revised | Yes | | Brain write confirmed | Exact addition (including bulk Extract-mode writes) shown and confirmed before writing | Yes | | Cold start handled | With no brain file the skill proceeded without redirecting; the only file created held Section 6 alone (or the block was shown for pasting) | Yes | | Learning Close ran | `/context/skill-sessions.md` has a new `type: execution` row for this session | Yes | ---
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.