corpus-retrieval

Corpus Retrieval

Overview

Iconic-startup PMF paths are the most credible counter to founder rationalization. When a founder claims "we don't need to refuse non-ICP revenue," surfacing DoorDash's refusal of non-restaurant merchants is more persuasive than abstract argument.

This skill indexes case files in skills/corpus-retrieval/cases/ by:

• Archetype (Hair on Fire / Hard Fact / Future Vision)
• Question resolved (Q1 / Q2 / Q3 / Q4, with timing if re-asked)
• Decision shape (e.g., wedge-narrowing, ICP-refinement, pricing-pivot, archetype-re-evaluation)

When invoked, returns 1-3 closest matches with the founder's verbatim quote where available + a route back to the relevant Q-skill or refuse-below-threshold gate.

The Iron Rule

NO FABRICATED CASES. NO PARAPHRASED-AS-VERBATIM QUOTES.

Every case in the corpus has a named source citation. If a case has no primary-source attribution (founder essay, Sequoia article, public interview, S-1 filing), it doesn't ship. Speculation about how Stripe found PMF is not evidence.

Case file schema

Each case file in cases/ follows this schema:

---
case_id: <kebab-case-unique>
founder: <named, multiple if co-founders>
company: <company-name>
archetype: <hair-on-fire | hard-fact | future-vision>
question_resolved: Q<N>
year_resolved: <year or range>
wrong_initial_hypothesis: "<verbatim or paraphrased with attribution>"
cold_outreach_count: "<verbatim from source>"
discovery_moment: "<what changed>"
what_they_refused: "<the off-the-table decision>"
pmf_signal: "<the rabid-demand evidence>"
decision_shape: <wedge-narrowing | icp-refinement | pricing-pivot | archetype-re-evaluation | other>
source: "<primary-source citation>"
confidence: <high | medium | low>
---

# <Company> — <one-line summary>

<2-4 paragraph narrative connecting Q1/Q2/Q3/Q4 resolution to the founder's
specific moves. Cite primary source for every quote.>

## What the founder did
- <bulleted move 1>
- <bulleted move 2>

## What they refused
- <bulleted refusal 1>
- <bulleted refusal 2>

## How this applies to your situation
<1 paragraph mapping the case's lesson to the founder's current question.>

v0.1.0 corpus inventory

5 cases shipped in v0.1.0, all sourced from Sequoia Arc framework article 2 (https://sequoiacap.com/article/pmf-framework-2/):

case_id	Company	Founders	Archetype	Q resolved	Decision shape
nubank-q1-q2-credit-cards	Nubank	David Vélez	hard-fact	Q1 + Q2	wedge-narrowing
doordash-q2-restaurant-icp	DoorDash	Tony Xu, Stanley Tang, Andy Fang, Evan Moore	hard-fact	Q2	icp-refinement
dropbox-q3-viral-onboarding	Dropbox	Drew Houston	hard-fact	Q3	retention-mechanism
square-q4-transaction-fees	Square/Block	Jack Dorsey	hard-fact	Q4	pricing-pivot
robinhood-q2-active-traders	Robinhood	Vlad Tenev	hair-on-fire (after 2022)	Q2 (re-asked)	archetype-re-evaluation + icp-refinement

v0.2.0 expansion target: 15-25 cases. Sources expand to First Round Review, NfX, founder retrospectives (Drew Houston's "The Dropbox Story", Patrick Collison's interviews), and primary-source S-1 filings.

Retrieval logic

When invoked with a query, match in this priority:

1. Exact decision-shape match (e.g., founder is doing wedge-narrowing → return Nubank).
2. Same archetype + same question (e.g., founder is Hard Fact at Q2 → return DoorDash).
3. Same question, different archetype (e.g., founder is Hair on Fire at Q2 → return Robinhood's 2022 re-evaluation).
4. Founder explicitly asks about a named company — return that case if present, else acknowledge absence.

Return 1-3 cases. More than 3 dilutes the signal.

Output format

Corpus matches (N=<count>):

1. <company> — <one-line summary>
   Archetype: <archetype> | Q resolved: Q<N> | Year: <year>
   Founder's verbatim move (where attributed): "<quote>"
   What they refused: <refusal>
   Why this applies to your situation: <one paragraph>
   Source: <citation>

2. <next case if applicable>

3. <next case if applicable>

Route back to: <arc-q<N>-<name> | refuse-below-threshold | founder-rationalization-defense>

When NOT to surface corpus matches

Don't surface cases:

• When the founder hasn't yet resolved archetype (route to archetype-detection first — cases are calibrated by archetype).
• When the founder's question is too vague for shape-matching ("how did Stripe do it?" without specifying which question).
• When the closest case is below confidence threshold (e.g., a case with confidence: low and no primary-source quote).

Surfacing weak cases dilutes the high-confidence ones.

Adding new cases

Per CLAUDE.md contributor guidelines:

1. Find a primary source (founder essay, named interview, Sequoia/First Round/NfX article, S-1 filing).
2. Extract the case using the schema. Quote verbatim where possible.
3. Set confidence: high only if all four schema fields (wrong_initial_hypothesis, cold_outreach_count, discovery_moment, what_they_refused) come from primary source.
4. Submit PR. The maintainer (Paul / EnactSkill) will pressure-test against the source.

No paraphrased-as-verbatim quotes. No "the founder probably did X" entries. Speculation is not the product.

Cross-references

• archetype-detection (prerequisite for accurate matching).
• arc-q1, arc-q2, arc-q3, arc-q4 (the questions the cases map to).
• refuse-below-threshold (corpus cases are evidence that thresholds matter).
• founder-rationalization-defense (corpus cases counter specific rationalizations).

---

v0.1.0: 5 cases from Sequoia article 2. v0.2.0: 15-25 case target with EnactSkill-generated case files from founder retrospectives.