market-corpus

Market Corpus

Overview

Market intelligence splits into two phases: corpus (facts, collected once, dated, expensive) and lens (analysis runs over the corpus, cheap, repeatable). This skill builds the corpus. Analysis is a separate skill: REQUIRED FOLLOW-ON: market-lens.

Core principle: no claim without a live source. Player names, funding amounts, pricing, headcount — anything from training data is a hypothesis, not a fact. Every fact in the corpus carries a URL, a confidence label, and a compiled date.

The Gates

Work moves through human-approved gates. Show the Intelligence Ledger at every check-in:

INTELLIGENCE LEDGER ──────────────── run N of market "<slug>"
✔ GATE 1  Roster       N firms (X auto + Y manual)   [approved <date>]
▶ GATE 2  Corpus       12/23 dossiers, evidence dated <range>
░ GATE 3+ Analysis     → market-lens skill

Each gate line states what's being captured and what question it answers.

GATE 1 — Roster

1. Define inclusion criteria with the user (segment, geography, stage, must-do/must-not-do).
2. Discover candidates via live search ONLY: Exa MCP (web_search_exa, company_research_exa), WebFetch on directories/funding news. Names you "already know" must be re-verified live — companies die, pivot, get acquired.
3. Present roster as a table (name, domain, 1-line why-included, discovery source) for human approval. Do not start Gate 2 without it.

GATE 2 — Corpus

For each approved company, build a dossier directory (layout + schemas in templates.md, located in this skill's directory):

<market-dir>/companies/<slug>/
  lens-facts.md  lens-services.md  lens-people.md  lens-jobs.md
  evidence.json  dossier.md

• One collector agent per company per lens file. Collectors gather facts and verbatim quotes only — no comparisons, no recommendations, no editorializing.
• Fan out collectors in batches of 4 companies max (rate limits, reviewability).
• Every claim gets a confidence label: verified (2+ independent sources or authoritative primary) / single-source / inferred (deduction, explained). Conflicts resolved toward primary sources; unresolved conflicts flagged inline, never silently picked.
• Every artifact header carries Compiled: . Staleness must be visible.
• One editor agent per company compiles the four lens files into dossier.md and extracts evidence.json (claim, section, source_url, confidence).

Gate 2 closes when all dossiers exist and the user approves the ledger. Then hand off to market-lens.

Tooling

Need	Tool
Web/company discovery	Exa MCP: web_search_exa, company_research_exa, people_search_exa
Page-level facts	WebFetch (pricing pages, job boards, press releases)
Funding cross-check	Crunchbase/PR via Exa; SEC EDGAR full-text search

Red Flags — STOP

• Writing a player list without a live search behind each name
• A funding figure, price, or market size with no URL next to it
• "Sources would include..." appended after the content — sources come per-claim, during collection
• Starting analysis/whitespace talk before Gate 2 is approved
• "User needs it fast, so skip the gates" — a fast wrong roster costs a rework; gates take minutes

Excuse	Reality
"I know this market"	Training data is stale. Companies on your list may be dead or acquired. Verify live.
"Deadline today, no time for gates"	Gate 1 approval is one message. Skipping it risks researching the wrong roster.
"I'll add sources at the end"	Untracked claims can't be sourced later. Capture URL at collection time.