Imported from alexleun/Human_in_the_loop_deep_research_skill (stock-analysis/SKILL.md). Install upstream with npx skills add alexleun/Human_in_the_loop_deep_research_skill --skill stock-analysis. Copyright stays with the author.

Stock Analysis Skill (v2.0)

A merged deep-research + equity-research workflow for analyzing an individual listed company — optimized for Hong Kong / HKEX listings (e.g. 1810.HK, 0700.HK), using only free, public, no-API-key sources. Produces an institutional-grade analyst note culminating in an explicit BUY / HOLD / SELL / AVOID / WATCH recommendation with a price target, confidence, conviction, and falsification criteria.

v2.0 changes (this is now the single canonical single-stock skill): Absorbed the institutional methodology formerly kept in stock-deep-research: adversarial Bull/Bear debate (06-debate.md), fragility audit & red-team (07-fragility-audit.md), deterministic governance gates + 7-state rating (08-synthesize.md), Goldman-style analyst note (09-report.md), and the SOTP/reverse-DCF requirements. Added a documented document-intelligence layer (search protocol, coverage report, deep-read appraisals, evidence #access/#evidence/#status tags, documents/findings-index.json, event timeline). The stock-deep-research skill is now the whole-market (index-level) skill — see its README row.

v1.1 changes (record): Added explicit "STOP and ask" human-approval gates (Principle 13), mandatory fetch-and-save source preservation with an end-condition compliance check (Phase 2), a technical-data fallback clause (Phase 5), DCF sensitivity-before-compute and real-peer-table discipline (Phase 4), a multi-stock file layout, post-report lifecycle, phase-revisit triggers, and an explicit done-state checklist. Based on skill-evolution lessons from the 0066.HK and 1810.HK analysis runs (2026-09-04).

v1.2 changes (record): Added an environment probe (Phase 1); primary-filing PDF extraction fallback with logged substitution (Phase 2); earnings reconciliation + per-source reporting-currency discipline (Phase 3); a units/currency sanity gate, required reverse-DCF, forward-EPS dispersion check, and peer FX rule (Phase 4); a price-freshness check and standard price+MA+volume chart output (Phase 5); a numeric confidence-scoring rubric (Phase 8); timestamped report filenames + FX/artifacts references (Phase 9); a task-ticking discipline; and a layout sync (price-history CSV, report/charts/, mandatory evolution log). Based on the stock-analysis v1.1 run on 0992.HK (Lenovo, 2026-09-05).

The final investment call always belongs to the human. This skill proposes; the user disposes.

This skill adapts best-practice patterns from the public literature:

When to Use This Skill

• Reaching a written BUY/HOLD/SELL view on a single listed company, with the rigor of an equity research note.
• Target is a HKEX / SEHK stock (though the flow generalizes to US/CN/JP/KR).
• Data must come from free, public sources — no paid API key.
• Output is a detailed research report + recommendation, not a trading bot or live dashboard.

Not for: analyzing a whole market or index (use stock-deep-research), executing trades, portfolio rebalancing, backtesting strategies, or building a factor model.

Core Principles

P1. Evidence Before Narrative (anti-hallucination)

Narrative must be translated into economics and every factual output must carry an exact quote or citation (URL / source ID) and an as-of date. Use the Fact / Derived-Fact / Analysis taxonomy:

• Fact — directly supported by a cited or frozen source.
• Derived fact — calculated from supported inputs with an explicit formula and matching time boundary (e.g. market cap from price × shares).
• Analysis — interpretation, scenario, or valuation-model conclusion (labeled as such).

A model may complete an analysis chain; it may never invent a missing fact. Any zero-source number is stripped from the report.

P2. Public-Source Only (No API Keys)

All data from free, public sources. If a source is paywalled/API-gated, substitute a free alternative and document it. Never fabricate data.

P3. Code-First Computation

All ratios, CAGRs, DCF, reverse-DCF, SOTP, and scenario math are computed by running Python on collected figures — never from LLM memory. On Windows, prefix scripts with sys.stdout.reconfigure(encoding='utf-8') (try/except wrapped) to avoid cp950 console crashes.

P4. Blended Analysis (Fundamental + Valuation + Technical + Debate)

A defensible call requires all of:

• Fundamentals — quality, growth, profitability, financial health
• Valuation — relative + intrinsic (fair-value range + price target), including DCF, reverse-DCF (what is the market pricing in?), and SOTP where the company has mixed-quality businesses
• Technicals & regime — trend, momentum, support/resistance, market regime
• Adversarial debate — the Bull and Bear cases argued at their strongest, plus catalysts and risks

P5. Adversarial Bull/Bear Debate

Run the Bull and Bear cases independently and at their strongest. Do not let one side see the other's conclusion first. The synthesis weighs both strongest cases rather than a one-sided narrative. Hallucinations from one side get refuted by the other before reaching the final call.

P6. Governance Gates (deterministic guardrails)

Before publishing a rating, run rule-based, non-negotiable gates that cannot be argued out of position:

• Quality Gate — refuse BUY on low-quality companies (no profitability, weak balance sheet).
• Regime Gate — if both technical and fundamental are bearish, demote BUY → HOLD.
• Sanity Override — if all input signals point one way but the call says the opposite, flag inconsistency and demote.
• Critical-News Override — CRITICAL negative event (fraud / bankruptcy / sanctions / major litigation) forces SELL/AVOID regardless of a cheap valuation.

P7. Integrity of Recommendation

Give a rating, price target, horizon, confidence (0–1), and conviction (HIGH/MEDIUM/LOW). If data is thin or contradictory, say so. Distinguish SELL (hold the position to exit) from AVOID (material risk — do not engage). Use WATCH when conviction is low but no material risk. Always end with a risk/disclaimer statement: research, not personalized advice.

P8. Falsification Discipline

Every thesis must carry an explicit falsification criterion and action conditions (monitoring signals that confirm/invalidate the view). If the criterion triggers, the recommendation is revisited immediately.

P9. Anti-Lookahead / Point-in-Time Discipline

Distinguish pre-cutoff evidence (findable facts) from post-cutoff reasoning (outcome anticipation). Do not present future knowledge as if it were known at decision time. Never use data dated after the "as-of" decision date to justify the decision.

P10. Fragility Audit (not just a disclaimer)

Treat fragility as a valuation input, not an appendix. Assess geographic/channel/customer concentration, policy dependence, litigation/IP exposure, supply-chain bottlenecks, inventory/warranty risk. Decide whether each risk warrants disclosure only, a multiple haircut, or a scenario discount.

P11. Source Preservation (P20 deep-research)

Save a local copy of every source before extracting findings (sources/…, documents/…) and index them in sources_index.md. Without local copies, findings are unrecoverable if URLs change.

P12. End-Conditions Discipline

Every phase defines end conditions as a checklist. The LLM exits a phase only when all end conditions are true, and the human verifies at gates.

P13. Human-Approval Gates (explicit "STOP and ask")

• Gate 1 (Phase 1 → 2): STOP — present the proposed scope, horizon, investor profile, and output contract; do NOT proceed to Phase 2 until the human confirms.
• Gate 2 (Phase 8 → 9): STOP — present the proposed recommendation (rating, target, confidence, conviction, thesis) produced in Phase 8; do NOT proceed to the analyst note (Phase 9) until the human approves.
• Gate 3 (Phase 9): STOP — present the final analyst note; do NOT mark the analysis complete until the human accepts it. Gates are implemented as explicit "stop and ask" mechanisms, not conceptual checkpoints. The LLM proposes, the human disposes.

Workflow Overview

Phase 1: Scope ── stock, exchange/code, horizon, investor profile, output contract
   │  (human approval gate 1)
   ▼
Phase 2: Collect & Preserve ── public sources + document-intelligence intake (corpus, coverage report)
   │
   ▼
Phase 3: Fundamentals ── quality, growth, profitability, financial health + deep-read appraisals
   │
   ▼
Phase 4: Valuation ── relative + DCF + reverse-DCF + SOTP → fair value range + price target
   │
   ▼
Phase 5: Technical & Regime ── trend, momentum, support/resistance, market regime
   │
   ▼
Phase 6: Bull/Bear Debate & Catalysts ── adversarial cases, catalysts, risks, event timeline
   │
   ▼
Phase 7: Fragility Audit & Red-Team ── concentration, contested signals, anti-lookahead, counter-hypotheses
   │
   ▼
Phase 8: Synthesize & Govern ── governance gates → rating + target + confidence + conviction
   │  (human approval gate 2)
   ▼
Phase 9: Analyst Note ── narrative → fundamentals → valuation → fragility → conclusion + action conditions
   │  (human approval gate 3)

Phase revisiting is normal. A fragility surprise in Phase 7 may send you back to Phases 3–4. Do not force a single pass.

Phase revisiting triggers (proactive, not reactive): When any of these fire, loop back to the referenced phase and update the affected computations before continuing:

• A new source or announcement changes a core financial input → revisit Phase 2–3 and recompute the affected metrics.
• Phase 4 valuation output contradicts a Phase 3 growth assumption → revisit Phase 3 and update growth before finalizing valuation.
• Phase 6 reveals a risk that changes a growth/valuation assumption → revisit Phase 3–4 before synthesis.
• Phase 7 fragility audit finds a concentration/anti-lookahead issue that invalidates an earlier input → revisit the affected phase.

Each revisit must be recorded in task_state.json (updated phases + reason) before moving on.

Phase Router

Before each phase, load its phase file with the read tool — e.g. before Phase 4, read 04-valuation.md and follow its procedure and end conditions. Do not work from this overview alone.

Suggested File Layout

{project_root}/
├── stock-analysis-{code}/       # one suffixed dir per analysis (multi-stock support)
│   ├── task_state.json          # multi-session state (phase, progress, decisions, environment line)
│   ├── sources/
│   │   ├── sources_index.md     # every source: URL + local copy + access date + evidence_status/access tags
│   │   └── YYYY-MM-DD-<desc>.{html,pdf,md}   # preserved local copies (mandatory)
│   ├── documents/               # document-intelligence layer
│   │   ├── search-protocol.md   # written before searching (types, hierarchy, criteria)
│   │   ├── coverage-report.md   # honest gap flagging (never padded)
│   │   ├── candidate-list.md    # deduplicated candidate documents
│   │   ├── appraisals/          # {doc-id}.md deep-read appraisals (verbatim quotes)
│   │   ├── findings-index.json  # machine-readable finding index (id, source_doc, quote, tags)
│   │   └── event-timeline.md    # dated market/company events (anti-lookahead cut-off)
│   ├── data/                    # extracted figures / datasets
│   │   └── <code>-price-history-<asof>.csv    # daily price series (technical baseline)
│   ├── notebooks/               # Python computation scripts (fundamentals, valuation, technicals)
│   ├── debate/                  # bull.md, bear.md, judge_notes.md
│   ├── report/
│   │   ├── <Code>-<Company>-<YYYY-MM-DD>-analysis.md   # timestamped final report
│   │   └── charts/
│   │       └── <code>-price-<asof>.png                # price + MA + volume chart
│   └── skill-evolution-log.md   # lessons learned (mandatory)

Task Ticking Discipline

Running inside an OpenSpec change, tick tasks at the end of each phase, not at archive time. When a phase's end conditions pass, mark its - [ ] rows complete in tasks.md and sync task_state.json (current_phase, status, artifacts) before starting the next phase. Write/fill the change artifacts (scope, specs, design) as the phases run — an archive written in one pass at the end hides where each phase actually finished. A single end-of-run sweep is not acceptable.

HK-Specific Data Notes

• Company announcements: always cross-check the HKEXnews portal (www1.hkexnews.hk) — the authoritative, free filing source for HKEX-listed companies. Record once: HKEXnews annual/interim PDFs are generally NOT machine-extractable via text fetch — record the limitation once here (and in sources_index.md) and go straight to the documented access ladder (aggregator substitute, logged) on subsequent runs; do not re-attempt extraction per submission.
• Reporting currency: HK companies report in HKD unless stated otherwise; some (yuan- or USD-reporting) need a consistent-currency note when comparing.
• Price sources: free aggregators (Yahoo Finance HK, AAStocks, Investing.com) may differ; pick one primary source, state it, and record "as of" timestamps.
• Liquidity/halts: HK stocks can have suspensions and a closing-auction regime — note any halt affecting the technical picture.
• Northbound/connectivity (Stock Connect) and HK/China policy can be material catalysts and risks — include them in the debate phase.

Post-Report Lifecycle

The analysis does not end at note approval. After Gate 3:

1. Archive the OpenSpec change (if the analysis ran inside one): openspec archive <change-name> --yes.
2. Preserve the spec for future diffing — the archived change keeps the baseline (scope, specs, design) for comparison.
3. Set a review reminder — explicit date AND reason (e.g. {date: 2026-09-13, reason: weekly recheck of 74.40 breakout / 65 support}), recorded in task_state.json.review_reminder.
4. Start a review with openspec-new-change using the same scope, and diff the new findings against the archived baseline. The note's action conditions and falsification criteria tell you what to re-check.
5. Update the skill-evolution log with what worked / what caused problems for the next analysis.

Completion Checklist ("Done" State)

The analysis is complete — nothing more to add — when ALL of the following are true:

1. ✅ Analyst note written and saved at a timestamped path report/<Code>-<Company>-<YYYY-MM-DD>-analysis.md
2. ✅ task_state.json updated with the final decision (rating, target, confidence score, status)
3. ✅ Human has approved the final note (Gate 3)
4. ✅ OpenSpec change archived (if run inside one), with tasks.md fully ticked
5. ✅ Review reminder set in task_state.json with an explicit date AND reason
6. ✅ Skill evolution log updated (mandatory)

Guardrails

• Evidence before narrative — every number needs a citation and as-of date (P1)
• Propose before deciding — no rating without all prior phases complete (P13)
• Code, don't memorize — compute metrics by script (P3)
• Debate both sides at their strongest — no one-sided narrative (P5)
• Gates are non-negotiable — Quality / Regime / Sanity / Critical-News overrides (P6)
• Fragility is a valuation input, not a disclaimer (P10)
• Log and preserve every source — local copy mandatory (P11)
• Check end-conditions before leaving any phase (P12)
• No lookahead — pre-cutoff facts vs post-cutoff reasoning (P9)
• Verbatim quotes only — no paraphrased findings without a quote; no fabricated quotes (P1, document-intelligence)
• State confidence, conviction, and falsification criteria — honest ratings only (P7, P8)
• Human disposes — user makes the final call (P13)
• Sanity-check script outputs — units/currency assert before a number enters the note (Phase 4)
• Reconcile earnings — never carry a reported-vs-segment-vs-adjusted contradiction silently (Phase 3)
• Check source freshness — exclude stale price sources (Phase 5)
• Canonical file set — keep exactly one canonical phase-file set; a superseded file (e.g. 06-catalysts-risks.md, 07-synthesize.md, 08-report.md) is renamed *.legacy.md, never left ambiguous beside the canonical file
• Keep spec requirement headers stable — when updating main specs, keep requirement headers identical and only change the body (a header rename breaks openspec archive); plan any rename explicitly