market-lens

Market Lens

Overview

Lenses are analysis runs over a corpus built by the market-corpus skill. A lens is cheap and repeatable; the corpus is the expensive part. Description before synthesis, always — you cannot say where the whitespace is until the descriptive map of what exists is approved.

REQUIRED BACKGROUND: market-corpus (gates 1–2, confidence labels, evidence.json format).

Preconditions — check before any analysis

1. Corpus exists at <market-dir>/companies/*/dossier.md + evidence.json. No corpus → STOP, run market-corpus first. Never analyze from memory.
2. Staleness check. Read Compiled: dates. If evidence is >30 days old, surface it in the ledger and ask the user: refresh or proceed-with-flag. Never silently analyze stale data.
3. Show the Intelligence Ledger:

INTELLIGENCE LEDGER ──────────────── run N of market "<slug>"
✔ GATE 1  Roster        23 firms                     [approved Jun 6]
✔ GATE 2  Corpus        23/23 dossiers, evidence Jun 6–8
▶ GATE 3  Descriptive   pricing ▣  packaging ▣  gtm ░░ 60%
          └─ capturing: who publishes prices, what's gated, why
░ GATE 4  Synthesis     BLOCKED until Gate 3 approved
          └─ will answer: "where is the pricing whitespace for us?"

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

GATE 3 — Descriptive lens runs

Pick lens packs for the user's question (definitions in lens-packs.md, located in this skill's directory). For each lens:

• One editorial author per lens — never a swarm. Extraction over many dossiers may fan out (lens runs are small; fan out freely), but exactly ONE agent writes the lens analysis document. Multiple writers merged together is how AI slop happens.
• Output to <market-dir>/analysis/dimensions/<lens>.md, header carrying compiled date + corpus date range.
• Descriptive only: what IS (who publishes prices, what's gated, patterns observed). No recommendations, no whitespace claims yet.
• Confidence labels travel. Every cross-company claim cites the dossiers/evidence it came from and inherits the weakest underlying confidence. A pattern built on inferred claims is inferred.

Present Gate 3 outputs for human approval before synthesis.

GATE 4 — Synthesis

Only after Gate 3 approval. One author. Answers the user's actual question ("where is the whitespace for us?") by reasoning ONLY over approved descriptive layers — if a synthesis point needs a fact not in the corpus, flag it as a corpus gap, don't fill it from general knowledge.

Deliverable: consolidated, structured, with per-claim source trail and an explicit Confidence & Gaps section (what's inferred, what's stale, what's unknown). End with Decisions needed / Next steps.

Red Flags — STOP

• Answering a market question with no corpus behind it ("I'll just do a quick take")
• Whitespace/recommendation language appearing in a Gate 3 document
• A cross-company claim with no dossier citation
• Synthesis written by merging multiple agents' editorial outputs
• Skipping the staleness check because "it was built recently" — check the dates, don't assume

Excuse	Reality
"The user asked for the answer, not a process"	Premature synthesis caused a full positioning rework once. Gates are the answer arriving correct.
"Corpus is probably fresh"	Reading dates takes seconds. Stale evidence silently poisons synthesis.
"More agents = better synthesis"	Editorial swarms produce homogenized slop. One author, reviewed, wins.