Company Skills

中文文档：README.zh-CN.md

A public-evidence company research toolkit with two complementary agent skills:

Both skills start from business economics and profit pools rather than titles, sentiment, or salary-site anecdotes. They separate public facts, user-provided facts, assumptions, derived metrics, and inference.

The toolkit runs in Claude Code, Codex, and compatible skill hosts. The talent-economics skill supports multilingual typesetting; the layoff-risk skill uses an adaptive candidate interview plus current public research.

A directory of auditable artifacts, with report.pdf as the deliverable:

The method runs one causal chain, in this order:

business model → profit pools → critical processes → critical roles
→ org/level structure → market replacement cost → compensation distribution → promotion paths

A title is not a salary. Two people at the same nominal level are worth different amounts if they control different profit pools, customer relationships or technical bottlenecks — so the analysis starts from where the money is made, not from a salary survey.

Pick one. The first is the shortest; the second is the most portable.

/plugin marketplace add kelvinfkr/company_skill
/plugin install company-talent-economics@company-skill

git clone https://github.com/kelvinfkr/company_skill.git
cd company_skill
bash install.sh

The current installer symlinks company-talent-economics into ~/.claude/skills/ and ~/.codex/skills/, then checks its dependencies. Because it is a symlink, git pull updates both hosts at once.

Install estimate-layoff-risk separately:

ln -s "$(pwd)/skills/estimate-layoff-risk" ~/.codex/skills/estimate-layoff-risk
# or: cp -R skills/estimate-layoff-risk ~/.codex/skills/

For Claude Code, use ~/.claude/skills/estimate-layoff-risk as the destination instead.

bash install.sh --claude      # Claude Code only
bash install.sh --codex       # Codex only
bash install.sh --project     # also into ./.claude/skills of the current repo
bash install.sh --copy        # copy instead of symlink
bash install.sh --check       # verify dependencies, install nothing
bash install.sh --zip         # build dist/company-talent-economics.zip
bash install.sh --uninstall   # remove the links

One-liner, no clone needed (it fetches the repo into ~/.cache first):

curl -fsSL https://raw.githubusercontent.com/kelvinfkr/company_skill/main/install.sh | bash

bash install.sh --zip

Then upload dist/company-talent-economics.zip at Settings → Capabilities → Skills → Upload skill.

Copy or symlink skills/company-talent-economics/ into any of:

Everything except the renderer is standard library only. Without LibreOffice, render_report.py --docx-only still produces the DOCX. bash install.sh --check tells you exactly what is missing and what it affects.

Ask for compensation and talent economics:

小米集团的员工薪酬分布是怎样的？年薪超过 100 万的有多少人？
How many people at Siemens earn above €150k, and which roles justify it?
ソニーで年収2000万円を超える社員は何人くらいいますか？

Ask for involuntary job-loss risk:

Use estimate-layoff-risk to assess a Dreame product intern's 3/6/12-month risk.
Interview me first, then research current public evidence and produce scenario ranges.

Or invoke a skill directly:

See the layoff skill's README for its full interview prompt, profit-pool framework, privacy boundaries, CLI pipeline, and fictional worked example.

S=~/.claude/skills/company-talent-economics

python $S/scripts/init_run.py "小米集团" --out ./talent-economics/xiaomi --year 2025
python $S/scripts/check_phase.py ./talent-economics/xiaomi --next   # prints the next phase's steps
# ...do them...
python $S/scripts/check_phase.py ./talent-economics/xiaomi          # gate check; repeat to phase 10

init_run.py infers language, jurisdiction, currency and threshold from the company name and prints what it chose. Override any of them:

python $S/scripts/init_run.py "Nestlé S.A." --out ./tmp/nestle --year 2024 \
       --jurisdiction ch --language fr --currency CHF --threshold 250000

check_phase.py is the spine. Each gate mechanically refuses to advance until the work is real: sources are logged, every figure carries an evidence id, the modelled payroll reconciles with the disclosed payroll within ±10%, no TODO survives into report.json. The gate output is the to-do list.

The report follows the language of the company name. If you write your request in a different language from the company name, your language wins — you are the reader.

Resolution order: --language → $CTE_LANG → the script and legal form of the company name.

python $S/scripts/i18n.py detect "台積電股份有限公司"
# {"language": "zh-TW", "confidence": "high", "jurisdiction_hint": "tw", ...}

Two different things get called "language" here:

• prose language — the language the analysis is written in. Any language.
• chrome locale — section titles, table headers, claim-type labels, number grouping, fonts and text direction. Twelve shipped packs: en zh-CN zh-TW ja ko es fr de pt it ru ar

A language with no pack still produces a full report in that language; the structural labels fall back (zh-HK → zh-TW → en) and the renderer says so. Numbers are scaled the way each language actually says them — 1.2 亿, 120M, 1,2 Mrd., 2 億 — and Arabic and Hebrew render right-to-left, with mirrored tables and footer.

Add a language — one JSON file, no code:

python $S/scripts/check_locales.py --new nl   # scaffolds locales/nl.json from English
# translate the strings; keep the keys and every {placeholder}
python $S/scripts/check_locales.py            # verifies it

assets/jurisdictions.json carries, per country: filing venues to scope site: searches to, the local names of the documents worth fetching, the company registry, currency, a high-pay threshold prior, an employer social-contribution rate, and a note on what that regime is unusually good at — or unusually misleading about.

python $S/scripts/i18n.py jurisdictions            # list all
python $S/scripts/i18n.py jurisdictions japan      # one entry; aliases resolve

Aliases work, so us, usa, america, sec and nasdaq all reach the same entry. Add a country by appending an entry — again, no code change.

Search runs in two languages at once: filings are indexed under their local names (有価証券報告書, Vergütungsbericht, 사업보고서, Formulário de Referência), cross-border analyst coverage under English ones. Searching one side loses half the evidence.

Two registry fields are priors, not facts, and every override is recorded in the notes:

• threshold_default — the round local number that reads as "high earner" there (CNY 1,000,000, JPY 20,000,000, USD 250,000, INR 1 crore). Not a statistical cut-off; not comparable across countries at market FX. Your threshold always wins.
• employer_social_rate — roughly 1% in Denmark, roughly 32% in Spain. The wrong one silently breaks the payroll envelope, so a disclosed figure always replaces it.

These are enforced by the harness, not by good intentions:

• No private individual's pay is ever inferred. Public profiles support aggregate role taxonomy and career transitions, nothing more.
• Public sources only. No login walls, CAPTCHAs, robots restrictions or access controls, ever.
• A search snippet is never a source. The primary document gets fetched and read, or the number does not get used.
• Forum and crowdsourced salary data is Tier E — calibration only, never alone behind a headline number.
• No invented numbers. A figure that genuinely is not public gets declared: NOT AVAILABLE: <fact> - <what was searched>. For an unlisted company the run then continues marked provisional, the cover says so, and nothing in it may be presented as measured. It is a label for honest uncertainty, not a workaround.
• No false precision. Counts and shares come out as Monte Carlo intervals. Observed public-profile transition shares are never called promotion probabilities, because the true denominator is unknown.

.claude-plugin/            plugin + marketplace manifests
install.sh                 installer, dependency check, zip builder
skills/
  company-talent-economics/
    SKILL.md               compensation and career research workflow
    scripts/               ten-gate harness
    locales/               12 chrome locale packs
    assets/                jurisdiction registry, search terms, priors, templates
    references/            economics, evidence, localization, career methods
    schemas/               report / model / evidence JSON schemas
    examples/              worked runs and runnable fixtures
  estimate-layoff-risk/
    README.md              interaction guide and local run instructions
    SKILL.md               layoff-risk workflow and safety boundaries
    scripts/               case setup, interview, search, estimation, rendering, validation
    references/            profit pools, roles, leadership regimes, evidence, calibration
    schemas/               case / evidence / model / report JSON schemas
    examples/              fictional robotics product-intern case

python skills/company-talent-economics/scripts/check_locales.py    # packs + registry
python skills/company-talent-economics/scripts/selftest.py         # gates 0-10, DOCX only
python skills/company-talent-economics/scripts/selftest.py --pdf   # also convert to PDF

selftest.py runs three scenarios end to end — a fully disclosed listed company, an unlisted company that has to take the provisional path, and the literal-[ ]-in-prose case — then renders every locale. CI runs exactly these on every push.

MIT. See LICENSE.