skill
company-detective
AI company research skill that turns filings and management claims into source-linked detective case files.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/XiaoshengChen/company-detective itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026Company Detective — an open-source AI skill for company research
Company Detective turns company due diligence into a source-linked investigation game. It helps an AI agent analyze filings, earnings materials, management promises, and contradictory evidence without turning the result into a dry research memo.
What does Company Detective do?
It creates a case file with clue cards, a management-claim tracker, bull and bear interpretations, a jury verdict, confidence, and unresolved side quests. Every material conclusion must cite a clue ID and source URL.
Best for: public-company research, earnings review, management-claim tracking, competitive intelligence, and evidence-backed company analysis.
Not for: personalized investment advice, rumor amplification, or unsupported allegations.
Why use it?
Generic “analyze this company” prompts often produce fluent summaries with weak traceability. Company Detective adds a persistent evidence ledger, claim lifecycle states, counter-evidence, and a game layer that rewards good sources.
Install
Copy this repository into your agent's skills directory, then invoke $company-detective.
Use $company-detective to investigate whether Acme's margin-expansion story is supported by its last four quarters of primary evidence.
Input
- company and jurisdiction
- falsifiable investigation question
- time window
- primary or authoritative source material
- optional structured JSON matching
references/evidence-schema.md
Output
- source-linked clue deck
- management promise lifecycle tracker
- bull case, bear case, and jury view
- confidence and evidence-coverage audit
- unresolved questions worth researching next
Reproducible example
python3 scripts/build_casefile.py examples/acme_case.json --output /tmp/acme-case.md
python3 evals/evaluate.py
The example is synthetic. It demonstrates the format without presenting fictional data as facts about a real company.
Three-round autoresearch log
The evaluator and primary metric were frozen before iteration. Each round changed one capability group and was kept only when the score improved. See autoresearch/contract.md and autoresearch/results.tsv.
| Round | Single changed variable | Score | Decision |
|---|---|---|---|
| Baseline | Plain company evidence table | 20 | baseline |
| 1 | Stable clue IDs + claim lifecycle | 50 | keep |
| 2 | Bull/bear/jury counter-readings | 70 | keep |
| 3 | Evidence-rewarding game layer + uncertainty | 100 | keep |
Demand validation
Real Chinese social-search evidence showed active interest in AI-assisted financial-report analysis, but also skepticism that a “skill” adds value beyond a stronger model. That counter-evidence shaped the product: the durable value is the evidence ledger, claim states, and auditability—not a decorative prompt. See research/demand-evidence.md.
FAQ
Is this an investment recommendation tool?
No. It organizes evidence and competing interpretations. It does not issue personalized buy, sell, or hold advice.
Can it investigate private companies?
Yes, if the user supplies lawful, reliable evidence. Expect lower confidence when authoritative public records are unavailable.
Why is it game-like?
The game mechanics make long investigations easier to continue. Points come from source quality, corroboration, and contradiction resolution—not from dramatic accusations.
License
MIT