Nine-Chapter Equity Research Skill
Turns AI tools into institutional-grade equity analysts producing traceable-facts, reproducible-valuation, auditable-conclusion stock reports and earnings reviews.
SKILL.md explicitly forbids trade execution, requires sandboxed handling of user data, prompt-injection defense for fetched content (data only), 'not obtained' marking instead of memory-filling, and source+timestamp attribution; no sign of privilege overreach or covert exfiltration. Deducted: no explicit user-confirmation gate, connector/data-flow transparency depends on implementation, unverified publisher, no stated rollback/recovery.
Workflow is self-consistent (six steps, mandatory reads, checker P0/P1 loop); the repo has real CI (syntax checks, unittest, demo runs) and unit tests for check_research_output.py. Deducted: visible tests do not cover dcf.py key valuation paths; scripts only run when a Python-capable environment exists; failure-feedback quality on abnormal input unproven; static cap keeps this at 11.
Trigger definitions are precise (company name/ticker + research intent), non-fit ranges are explicit (one-line quotes, macro, portfolio allocation, HFT), bilingual output and A/H/ADR coverage fit Chinese users well. Deducted: the broad 'trigger on any company + research intent' risks false positives on casual questions; reliance on overseas filings/quote sources makes mainland-China network reachability unevaluated.
Well-layered docs (SKILL.md→references→industries→scripts), install notes, dependency table, MIT license, tag-driven versioned release workflow, sample reports. Deducted: no changelog file, maintenance ownership and update path rest implicitly on a single individual, sample data sources may not match user environments.
The design goal is differentiated and concrete (reproducible valuation, expectations-gap spine, earnings-quality veto), with specific declared output formats. Deducted: static review cannot verify the sample outputs are directly usable; 'deepest AI equity-research skill' is marketing language; output quality depends on data availability at run time; static cap keeps this below 7 at 6.
Three evidence types exist: real CI workflows, unit tests for the checker, and bilingual sample reports (NVDA/GOOGL), plus JSON-assumption filing for reproducible valuation. Deducted: sample data is not independently checkable, no test evidence for dcf.py numerical correctness, fact/judgment separation unverified in execution; static cap at 5.
- This is a static source review (low confidence); no scripts were executed and no sample reports reproduced.
- Visible unit tests cover the checker script only, not dcf.py valuation correctness; independently recompute key figures before relying on valuation output.
- The skill depends on fetching regulatory filings and quotes online; reachability of overseas data sources from mainland-China networks is unverified and may yield many 'not obtained' fields.
- Output is research reference only, not investment advice; verdict labels and position-sizing references should not drive trades directly.
- Broad trigger conditions may fire a lengthy deep-research pipeline on casual price questions, costing significant time and tokens.
- Publisher identity is unverified; maintenance continuity depends on a single individual author.
What does this skill do, and when should you use it?
This is an Agent Skill for single-stock research in public markets that converts requests like 'research NVDA' into a deep report following near-institutional discipline. It runs on an expectations-gap thesis (market-implied vs. independent expectations), performs forensic accounting checks (accruals, Beneish M-Score) before any valuation, and cross-validates with at least three valuation methods, all computed by script with assumptions archived as JSON. It offers a full mode (nine-chapter stock report) and an earnings mode (nine-chapter earnings deep-dive), covering US, Hong Kong and A-shares, including dual A/H listings and China VIE/ADR structures. Deliverable is a single PDF report with sources and timestamps.
Reads SKILL.md plus the templates, industry rules and data-sourcing manuals under references/, then runs a six-step workflow: after confirming the ticker and scope, it gathers primary disclosures, market/valuation anchors, consensus and industry data in parallel; reconciles key figures with source and timestamps and runs forensic accounting (accrual quality, M-Score, revenue-recognition red flags) to grade report credibility A–D; writes a nine-chapter report whose first chapter contains a decision triad and an expectations-gap table; runs valuations (reverse DCF+PVGO, probability-weighted three-scenario DCF, EPV, EVA, optional Monte Carlo) via scripts/dcf.py with assumptions saved as JSON and mental math forbidden; runs scripts/check_research_output.py consistency checks and a pre-mortem before finalizing; and delivers one report file (PDF by default, .md/.docx/.xlsx on request), keeping assumptions and script outputs as internal work files for auditability.
- An individual investor researching a company for the first time who wants a complete workfile of business, competition, financials and valuation rather than a news summary.
- A holder deciding 'what did this earnings report actually change?' after a quarterly/annual report or earnings call, using the earnings mode to dissect the expectations gap, guidance and valuation shifts.
- A long-term investor writing an investment memo who needs auditable, reviewable conclusions with prediction dates for verification.
- An investor covering dual-listed A/H names or China VIE/ADR structures who needs market-by-market comparisons and structural-risk pricing.
- An analyst in SaaS, semiconductors, banking or pharma who needs industry-specific KPIs (ARR, NRR, NIM, clinical probabilities) to reshape the model and valuation framework.
What are this skill's strengths and limitations?
- All valuations are computed by scripts/dcf.py with assumptions archived as JSON — mental math forbidden, conclusions reproducible.
- Forensic accounting (accruals, M-Score) precedes valuation; C/D credibility grades veto buy actions.
- Expectations-gap thesis enforces a falsifiable independent view — no buy/sell call without a divergence.
- Key figures require source + timestamp; conflicts must be reconciled and gaps honestly labeled 'not obtained'.
- 20 industry appendices and multi-market support (US/HK/A-shares, A/H, VIE/ADR) instead of one-size-fits-all templates.
- Depth-first design: heavy process and slow output; not for users wanting a one-line verdict.
- Web search/scraping is a recommended dependency; offline use requires the user to supply all material.
- PDF output depends on a pdf skill or an md-to-PDF toolchain and degrades to .md when unavailable.
- Only two example outputs (NVDA, GOOGL) are provided; there is no systematic test suite, and real-world behavior across platforms/connectors needs self-verification.
- Output is research reference only, not investment advice; decisions and consequences rest with the user.
How do you install this skill?
Simplest: paste the repo URL https://github.com/rollingSirius/equity-research-skill into any AI tool that supports skills and ask it to install. Claude Code: run git clone https://github.com/rollingSirius/equity-research-skill.git ~/.claude/skills/equity-research (user-level) or clone into the project's .claude/skills/ (project-level). Claude Desktop / Cowork: zip the repo and upload under Settings -> Capabilities -> Skills. Codex / other agents: place the repo in the project (e.g. skills/equity-research/) and instruct the agent to read and follow SKILL.md when asked to research a stock. Local Python is not a prerequisite; scripts (stdlib only) can run in the agent's code environment or an online notebook.
How do you use this skill?
Trigger in natural language, e.g. 'Research NVDA for me', 'Is Marvell a buy?', 'Deep-dive Apple's latest earnings', 'Analyze Tencent's latest results in earnings mode'. Or invoke explicitly: in Claude Code use /equity-research 分析 NVDA; for other agents say 'First read skills/equity-research/SKILL.md and follow its workflow, then research <stock>'. Output defaults to PDF; specify .md/.docx/.xlsx (valuation workbook) if preferred. Not suited to one-line price calls or high-frequency trading signals.