Finance & Investment Banking value-investingcompany-researchlong-form-writingfact-checkingfinancial-analysismulti-agentchinese-market

Deep Company Series (看懂XX)

A value-investing research skill that dissects one company in 3-8 long-form articles, from mindset reset to decision framework, with strict fact-checking standards.

FollowSkills review · FSRS-2.0
Not recommended
49/ 100 5-point scale 2.5 / 5
1 2 3 4 5 6
1Trust11 / 25 · 2.2/5

The skill explicitly requires public reports to use only public information and mandates grep scans for local usernames, /Users/, and personal identifiers before publishing, showing strong privacy awareness. However, the workflow relies on Bash commands and local file operations without explicit least-privilege or user confirmation steps; external network dependencies (Futu, broker research) pose uncontrollable data source risks. No malicious intent detected, but permission and confirmation mechanisms are incomplete, hence deductions.

2Reliability7 / 20 · 1.8/5

Provides detailed execution steps and checklists, but no test cases or CI for the skill's key paths. Scripts like financial_rigor.py have tests, but the skill itself does not reference them. Static reading cannot confirm stable operation; error handling and feedback lack evidence.

3Adaptability12 / 15 · 4.0/5

Clear scenario (for WeChat article writers wanting deep company series), with flexible article count adaptation, explicit templates, and banned words list, boundaries clear. Relies on overseas services (WebSearch, Futu, broker reports) which may be restricted in mainland China, but not fully dependent, so given high score with slight deduction.

4Convention9 / 15 · 3.0/5

Clear description, structured templates, concrete examples, and compliance requirements, but lacks versioning, changelog, and maintenance ownership details; depends on other repository files without full dependency list. MIT license clear, but provenance unverified.

5Effectiveness6 / 15 · 2.0/5

Clear goal (produce publishable deep long-form series), with guidance on writing style, fact-checking, revision workflow—theoretically helps users complete the task. However, no actual output examples or validation reports, cannot confirm output directly usable; producing long series requires significant user interaction, marginal value hard to assess statically.

6Verifiability4 / 10 · 2.0/5

Key claims (e.g., returns) from README lack independent verification evidence. No fabrication found, but no third-party corroboration or reproducible tests, hence low score.

Evidence confidence:Low Reviewed Aug 07, 2026 Reviewed revision 4ddc638fd536
The upstream repository has new commits since this review. The score still applies to the reviewed revision shown and may not cover the latest changes.
Before you use it
  • Skill relies on overseas network services (web search, Futu, etc.), which may be restricted in mainland China; test availability beforehand.
  • Skill requires user review of multiple long articles, high cost and time; assess actual needs.
  • Return data in README is personal claim, not independently verified; do not use as investment basis.
Review evidence [1][2][3][4][5][6]
See the full review method →

What does this skill do, and when should you use it?

This skill from the AI Berkshire repository generates the 'Understanding Company X' (看懂XX) long-form series: 3-8 standalone articles that together form a complete investment thesis on a single company. Its core IP is not 'writing well' but 'revising rigorously'—backed by a strict fact-checking checklist that eliminates pseudo-precision, subjective probability-weighted expectations, and absolutist language common in financial writing. It follows the methodologies of four value-investing masters (Buffett, Munger, Duan Yongping, Li Lu) and provides detailed article templates, style rules, and a revision workflow.

Given a company name ($ARGUMENTS), it produces a series of 3-8 articles (e.g., 8 articles for complex firms like Tencent, 3 for simpler ones like Kuaishou). Before research, it runs the date command to set the data cutoff. It requires reading the last 5 annual reports and at least 3 independent sell-side research reports. Each article follows a fixed skeleton: a header block, a hook with counter-intuitive numbers, the main body along one of 8 core axes (moat, profit engine, hidden assets, AI impact, financials, management, valuation, etc.), and a key-points recap at the end. Output goes to reports/{Company}/《看懂{Company}》/, or a date-suffixed folder if one exists. Before pushing, it runs grep to scan for privacy fields and uses python3 tools/financial_rigor.py for precise calculations and cross-validation.

  1. A financial content creator wants to write a deep-dive series on a company for a public WeChat channel, with each article standalone but sharing a consistent valuation framework.
  2. A value investor needs to dissect a complex company like Tencent (multi-business, hidden investments, rich management history) across 7-8 articles.
  3. An analyst covering a simpler company with a clear core business wants a minimal 3-article series without padding.
  4. A writer needs to revise a draft while ensuring cross-article consistency in numbers, terminology, and references.
  5. A user planning to publish a series publicly requires compliance and privacy checks, automatically scanning for personal identifiable information.

What are this skill's strengths and limitations?

Pros
  • Built-in rigorous fact-checking checklist prevents pseudo-precision, subjective weighted expectations, and absolutist claims
  • Flexible article count (3-8) adapts to company complexity
  • Emphasizes 'revising' over 'writing', with a defined severity hierarchy for revisions (hard errors, subjective, granular, unreliable)
  • Cross-article consistency scanning and cascade updates ensure numbers, terms, and references align
  • Clear compliance and privacy requirements—pre-push grep scan for sensitive data
Limitations
  • Heavy research workflow requiring 5 years of annual reports and at least 3 research reports; time-intensive
  • Relies on `/investment-research` or similar as internal drafts, which must be installed separately
  • No automated tests—only manual and grep checks, which may miss edge cases
  • Article template is derived from Tencent; simpler companies require manual merging of axes, adding learning overhead
  • Does not include direct examples for `financial_rigor.py`; users must reference the repo's tools folder

How do you install this skill?

This skill is located at codex-skills/deep-company-series/SKILL.md in the AI Berkshire repository. Clone the repo and install per your client: for Claude Code run ./scripts/install-claude-commands.sh (macOS/Linux) or .\scripts\install-claude-commands.bat (Windows); for Codex run ./scripts/install-codex-skills.sh/.bat. The repo also maintains source files under skills/*.md, and codex-skills are generated from them.

How do you use this skill?

In Claude Code, invoke /deep-company-series 腾讯 (or any company name). In Codex, restart after installation and describe the task, e.g., 'use deep-company-series for Tencent'. Once active, confirm the current date, then follow the workflow: Phase 1 (research), Phase 2 (write articles in order), Phase 3 (cross-article consistency scan), Phase 4 (final pre-publication checks). Adhere to the per-article structure template and the prohibited-words list, and strictly apply the fact-checking checklist.

How does this skill compare with similar options?

Within the AI Berkshire repo, /investment-research produces a single deep-dive report, while this skill focuses on a long-form series; /earnings-review centers on financial statement analysis, whereas this covers full company breakdown. Compared to asking an AI directly, this skill forces conclusions, uses multi-perspective adversarial thinking, and embeds structured anti-bias mechanisms to make output decision-ready.

FAQ

Is the number of articles fixed at 8?
No. Eight articles suit a high-complexity company like Tencent; simpler firms (e.g., Kuaishou) can be covered in 3. The rule: each core axis must be able to stand alone with one sharp question; if a draft starts repeating content, merge articles.
What if data sources conflict?
Prefer cross-verification: use shell/read to find original data, and present a three-way comparison (user's claim vs. what I find vs. what I previously used). If third-party estimates vary widely, deletion is more reliable than editing (unless the user insists).
How to manage token costs?
This skill performs multi-round research and is token-heavy. Control costs by adjusting workflow: use `/quality-screen` for initial screening, and apply this skill only to firms worth deeper research.

More skills from this repository

All from xbtlin/ai-berkshire

Related skills