Autonomous research prompts: the AI does the fieldwork, you review the evidence
Prompts in this directory follow the autonomous investigation
interaction mode (see interaction-modes.md).
Where /prompts/ assumes you hold the context and /prompt-generators/
facilitates it out of you, investigations point the AI at the world:
web search, published data, credible sources — with citations, labeled
inference, and defaults that let the prompt run without you.
That last property matters: because these prompts proceed on best-available evidence when nobody answers questions, they are suitable for agents, loops, and scheduled runs — a competitive scan that refreshes quarterly, a market snapshot that diffs against last month's.
- Question budget — a hard cap (usually 3), then proceed with labeled assumptions
- Search plan gate — a 3-bullet plan shown before researching; it continues unless you revise it
- Evidence labels — every key point marked Fact (source- supported), Inference (evidence-based interpretation), or Assumption (working guess)
- Real, checkable URLs — and a do-not-invent list naming the domain's specific fabrication risks
- Just Enough Mode by default — strongest findings, short bullets; Verbose only on request
- Stable output schema — so run N and run N+1 are diffable, which is what makes scheduled re-runs and delta reports possible
- Final Step block — exactly 4 numbered next options
full-spectrum-company-sweep-prompt.md runs all seven collection disciplines back to back on one company, fuses them with the confidence-stacking rule, and ends with a call-ready brief: a sixty-second summary, the three things worth saying, the questions you will be asked with answers, and the claims you must not make. One prompt, one run, no orchestration.
Use it when a company suddenly matters and you have one sitting to get smart. Everything below gives you more depth on a narrower question -- reach for those when this sweep tells you which channel is thin.
| Prompt | Best For |
|---|---|
| full-spectrum-company-sweep-prompt.md | Everything about one company in one run: all seven disciplines, fused, ending in a call-ready brief |
| competitive-research-snapshot-prompt.md | Just-enough competitive landscape research with cited snapshots, a comparison matrix, and "so what" implications |
| competitive-intel-watch-prompt.md | Scheduled delta monitoring against a prior snapshot: material shifts only, with battle-card update flags |
| market-landscape-scan-prompt.md | Mapping a market's segments, players, substitutes, and whitespace before sizing or positioning |
| voice-of-customer-miner-prompt.md | Mining reviews, app stores, and forums for unmet needs and competitor weaknesses, with quoted evidence |
| earnings-executive-signal-refresh-prompt.md | Quarterly diff of one company's executive and strategy signals against a prior profile |
| pricing-packaging-tracker-prompt.md | Tracking competitor pricing and packaging as a diffable time series |
| pestel-delta-monitor-prompt.md | Quarterly re-scan of macro factors: what moved, what broke, what entered the frame |
| tam-sam-som-analysis-prompt.md | Citation-backed market sizing, bottom-up, with best/base/worst sensitivity |
| battle-card-builder-prompt.md | Evidence-cited battle card from public sources: per-claim labels, field-action format |
| swot-analysis-prompt.md | SWOT with sources, quadrant discipline, and the S-O / W-T crossings |
| porters-five-forces-prompt.md | Industry structure read: five rated forces with documented signals, profit-pool implication |
| ansoff-matrix-prompt.md | Growth options with evidence per quadrant, risk gradient respected, recommended sequence |
| all-source-fusion-prompt.md | The situation room: fuses signals from the other investigations into confidence-rated stories with artifact-mapped responses |
Seven discipline sweeps, one per collection discipline in the compendium, each embedding that discipline's sources and signal → inference chains. All seven emit the same fusion-ready Signal Inventory schema, so any combination stacks in all-source-fusion-prompt.md and run N diffs against run N-1 on a schedule (see loops/fusion-cadence-routine.md for the governed rhythm).
| Prompt | Discipline | Primary feeds |
|---|---|---|
| osint-collection-prompt.md | Open Source: press, analysts, social, reviews, conferences, prediction markets | Battle cards, positioning |
| finint-collection-prompt.md | Financial: filings, earnings dodges, procurement, registrations, sovereign capital | Battle cards, SOM capture rates |
| geoint-demoint-collection-prompt.md | Geospatial and demographic: establishment counts, occupations, wages, trade flows | TAM/SAM/SOM, ICPs, personas, messaging |
| techint-collection-prompt.md | Technical: patents, trademarks, technographics, changelogs, standards, preprints | Roadmap bets |
| humint-collection-prompt.md | Human: hiring surges, talent moves, sentiment, win/loss framing | Roadmap bets, battle cards |
| sigint-collection-prompt.md | Signals: web/pricing diffs, SEO/SEM, SSL certs, app-store metadata | Battle cards, pricing strategy |
| masint-collection-prompt.md | Measurement and signature: supply chain, facilities, ops exhaust, certifications | Threat assessment, launch prediction |
The prompts above are the runnable layer. The doctrine behind them —
what to collect, where, and which signal → inference chains to run —
lives in reference/:
- competitive-research-compendium.md — the eight intelligence-community collection disciplines (OSINT, FININT, GEOINT/DEMOINT, TECHINT, HUMINT, SIGINT, MASINT, All-Source Fusion) mapped to PM artifacts, with sources, signal chains, confidence stacking, and a fusion cadence
- regional-source-overlays-eu-mena.md — EU and MENA source overlays: procurement platforms, registries, statistics bureaus, and the regional guardrails (the disciplines don't change across markets; the sources and evidentiary burden do)
Files in reference/ are doctrine, not prompts — they carry no
comment block and are excluded from the catalog and validator.
How this directory's prompts map onto the compendium's disciplines. Every discipline now has a dedicated collection sweep; the second column names the deep-dive prompts each sweep hands off to:
| Discipline | Collection sweep | Deep dives |
|---|---|---|
| OSINT | osint-collection | market-landscape-scan, competitive-research-snapshot, voice-of-customer-miner |
| FININT | finint-collection | earnings-executive-signal-refresh |
| GEOINT/DEMOINT | geoint-demoint-collection | tam-sam-som-analysis, pestel-delta-monitor |
| TECHINT | techint-collection | — |
| HUMINT | humint-collection | prompts/win-loss-analysis (synthesizes your team's interviews into ground truth) |
| SIGINT | sigint-collection | competitive-intel-watch, pricing-packaging-tracker |
| MASINT | masint-collection | — (anomalies disambiguate via FININT and HUMINT) |
| All-Source Fusion | all-source-fusion | swot-analysis, porters-five-forces, ansoff-matrix, and battle-card-builder consume the fused evidence |
Three layers, one motion — or full-spectrum-company-sweep-prompt.md, which walks all three internally for a single company when you do not have time to orchestrate.
- Collect — instantiate the engagement (the compendium's six variables: TARGET, MARKET, GEOGRAPHY, BUYER, CAPABILITY, DECISION), then run the collection sweeps for the disciplines that matter. New to a market? Start wider: landscape scan → snapshot the players → then sweep.
- Fuse — when 2+ disciplines hold signals, run all-source fusion: independence test, confidence stacking, conflict digs. Your team's win/loss interviews (prompts/win-loss-analysis-prompt.md) are the ground truth that confirms or refutes what the public signals inferred.
- Act — actionable stories update artifacts: five forces reads the industry's structure, SWOT one player's position, Ansoff the growth options, TAM/SAM/SOM the size, and the battle-card builder arms the field. Then schedule the rhythm with loops/fusion-cadence-routine.md.
Worked example — a launch caught before the press release: your
weekly SIGINT sweep finds a new SSL certificate for
analytics.competitor.com (launch staging, weeks ahead). That flags
TECHINT, which finds an 8-filing patent cluster in analytics
classifications plus two arXiv papers by their staff. TECHINT flags
its fusion pair, HUMINT, which finds 25 data-engineering postings
this quarter against a baseline of 4. Fusion stacks it: three
independent disciplines, one story — actionable. FININT's earnings
pass adds the CFO dodging an analytics capex question (intent
corroborated by money). Response before their launch: battle-card
maturity play, roadmap accelerate/concede call. Your next win/loss
round confirms whether analytics actually decides deals — and the
ledger feeds the next fusion run.
To build a custom investigation, use prompt-generators/research-agent-prompt-generator.md. To turn any output here into a stakeholder story, use storytelling/Generator - Research-to-Narrative Bridge.md. For loop, batch, and routine recipes over these prompts, see loops/ — swot-batch, market-sizing-loop, and competitive-watch-routine run this directory's frameworks under governed repetition, and fusion-cadence-routine runs the whole collection floor on the compendium's weekly / monthly / quarterly / annual rhythm.