deep-company-series

Deep-Company Series: An 8-Part Deep Dive on One Company

Write an 8-part deep-dive series (~120k words total) on a single company, from cognitive reset to a decision framework. The core IP is not "writing well" but "revising strictly" — most finance long-form violates this skill's fact-check standard.

1. When to Use

The user wants "textbook-level" deep research on a company, published as a series of long-form articles. Distinct from a single research report:

• 8 parts, ~120k words, full loop from cognitive reset to a decision framework
• Each part stands alone (shareable singly) but shares one valuation / management / price framework
• Written for readers willing to spend 90 minutes understanding one company

Not for: a single research report, earnings note, sector study — use other skills (fundamentals / earnings / sector).

2. Series Template (8 Parts)

Plus 00-series-overview.md as an index (unpublished).

3. Writing Style

Voice

• Direct, sharp, no filler — open with a number or a counterintuitive claim
• Value-investing frame — Buffett/Munger/Duan Yongping/Li Lu lenses woven in (no name-dropping)
• No preset stance — data first, logic next, conclusion last
• Show both sides — every core judgment carries a "but on the other hand..."
• Mobile preview — the first 18-20 characters must stand alone

Banned Words

Title Style

• Hook with a contrast number or a counter-consensus claim ("15 years, 7 failed challenges"; "salary 42.92M = 0.0017% of profit")
• Neutral subtitle summarizing content
• Avoid hype metaphors: "the next Buffett", "the X of China", "GOAT" — all banned

4. Strict Fact-Check Checklist (the Core IP)

"Pseudo-precision" traps to watch for before writing

1. Probability-weighted expected value: 30% × A + 50% × B + 20% × C = expected +X% is almost always garbage — the probabilities are pure subjective, giving readers false precision. List scenarios + triggers + direction only; do not compute a weighted expectation.
2. Third-party MAU/share estimates: QuestMobile / 七麦 / CBNData differ hugely (2-3× at the same point). Use only the two most-credible as anchors; describe the rest qualitatively.
3. Linear extrapolation of historical growth: 2025 +33% × 5y CAGR → 2030 X is financial illiteracy. Use scenario assumptions + high/low ranges; never a promise.
4. Undisclosed shareholding: unlisted-company stakes are never publicly disclosed. Give a range, mark "unknowable".
5. Strong attribution: "competitor failed because of X." List multiple causes; this article does no single attribution.

The 7 mandatory revision checks

□ 1. Cross-article number consistency: market cap, Non-IFRS net income, key holding % aligned across the series
□ 2. Caliber labeling: Non-IFRS / GAAP / Non-IFRS-SBC / FCF — which is used, clear throughout
□ 3. Double-counting scan: consolidated subs are NOT in the "investment portfolio"; SOTP doesn't count them twice
□ 4. Peer-comparison fairness: don't compare "core-business PE (cash + portfolio stripped)" with "peer PE (not stripped)"
□ 5. Probability-weighted expectations deleted (see above)
□ 6. Absolute language softened: grep "obviously|inevitably|severely|textbook|perfect"
□ 7. Third-party data sourced: every non-filing data point followed by "(source: X)"

Known hard-error risks (list before writing)

• Historical return multiples: use cumulative-invested basis (e.g. Riot 33×, not 58×)
• Shareholding %: use the latest filing/financial-app basis (e.g. Tencent's Meituan stake changes with disposals)
• "Distribution accounting": treated as disposal gain under IFRIC 17, recognized on declaration date
• Share count rebounds: SBC granted in clusters at year-start can lift share count short-term

5. Execution

Phase 1: Research (before writing 01-02)

1. get_financial_statements — last 5 years of annuals, latest quarterly
2. get_research_reports / web_search — at least 3 independent sell-side reports (find consensus + dissent)
3. Optional: run_swarm (e.g. equity_research_team or value_investing_committee) to generate an internal research draft
4. Confirm the 8-part core theses with the user (avoid writing the wrong direction)

Phase 2: Writing (01→08 in order, no skipping)

• After each part, write_file to reports/{company}/《Understanding {company}》/0X-XX.md
• Don't publish immediately — wait for user review
• Revise on feedback

Phase 3: Cross-Article Consistency Scan (after all 8)

This is the key differentiator. Use tools to scan:

1. read_file each part + report_audit (command=extract) to pull numbers (market cap, net income, holding %, PE) from each
2. Cross-check the same number across parts — use financial_rigor (command=cross_validate) to cross-validate the same metric's values across articles; flag >1% deviation as a caliber mismatch
3. read_file checks: is each term (FBS, SBC, Non-IFRS) defined at first use; do "see part 06" references actually resolve; do recaps match body numbers
4. Absolute-language scan: grep "obviously|inevitably|severely|perfect" and soften each

Phase 4: Pre-publish Final Check

• report_audit (command=verdict) as a gate on each part: extract numbers → verify → PASS/FAIL
• Confirm all numbers are traceable, no pseudo-precision, no absolutism

6. Revision-Feedback Handling

1. Verify facts first (don't just change)

If the user says "X is wrong", use get_financial_statements / web_search to cross-check the original; present "user's number vs what I found vs what I used".

2. Grade the revision

3. Cascade check after a fix

Before fixing one spot, think "where else is this number/concept referenced":

• Market cap changed → cascade to PE / core-business PE / discount / FCF yield
• Holding % changed → fix TOP-10 sort + historical holding table + disposal list
• Caliber changed → fix first definition + later references + recap

7. What This Skill Does NOT Do

• Does not make investment decisions for the reader — every part ends with "not investment advice"
• Does not predict prices — only "scenarios + triggers"
• Does not compute a weighted "expected annualized return" — subjective probability misleads
• Does not write "famous investor X also holds" — using someone else's holding to back your judgment is anti-value-investing
• Does not force all 8 parts — if a part lacks enough standalone content (e.g. management isn't distinctive), merge it or reduce the count

One-liner: writing an "Understanding X" series is about revising strictly, not writing well — most finance long-form dies from pseudo-precise numbers, subjective weighted expectations, and absolute language. This skill exists to flag all those traps before writing and sweep them clean after (report_audit + financial_rigor.cross_validate).