finance-skills 0.3.0

Analyst-style equity research over one shared engine: segment-aware Rule of 40, DCF, red flags, screening — read-only, not investment advice.

finance-skills

Analyst-style equity research as an agent skill — driven by real fundamentals, not narrated numbers. Ask about any public ticker in plain English and get a report built from a fetch → compute pipeline.

Read-only. Not investment advice. It only reads public market data, never places trades, and every figure should be verified against primary filings.

/finance-skills analyze Do you think NBIS is a buy?
/finance-skills is NVDA overvalued?
/finance-skills compare AMD and NVDA

What makes it different

Most "Rule of 40" tools compute one flat number. This one classifies the company's growth/capital regime first, then picks the right formula and peer benchmark — the mistake experienced analysts flag most often on names like CoreWeave (CRWV) and Nebius (NBIS), whose GPU capex breaks the classic formula. See references/rule40.md.

Install as a skill (/finance-skills)

The whole repo installs as one skill for Claude Code, Antigravity, or a Codex-compatible tool. Easiest — ask your agent:

Hey Claude, install this skill https://github.com/notEhEnG/finance-skills

Or one command:

curl -fsSL https://raw.githubusercontent.com/notEhEnG/finance-skills/main/install.sh | bash -s -- claude
# or: ... | bash -s -- antigravity   (or codex, or all)

Or manually copy the repo into the tool's skill dir (scripts/ must sit next to SKILL.md so the engine can run):

Tool	Install path	Invoke
Claude Code	.claude/skills/finance-skills/SKILL.md	/finance-skills analyze Do you think NBIS is a buy?
Antigravity IDE	.antigravity/skills/finance-skills/SKILL.md	/finance-skills is NVDA a buy? or assign to an agent
Codex-compatible	<codex-skill-dir>/finance-skills/SKILL.md	/finance-skills ...

Then, if needed: pip install yfinance.

How invocation works

/finance-skills [verb] <plain-English question>. The skill (1) extracts the ticker (mapping company names, e.g. Nebius → NBIS), (2) determines intent (an explicit verb or inferred from the question), (3) runs the engine and answers the actual question with cited numbers. So all of these are equivalent:

/finance-skills is NBIS a buy?
/finance-skills analyze Do you think NBIS is a buy?
/finance-skills analyze NBIS

Keyword routing (plain English → verb)

Step 2 isn't a guess. The router carries an explicit keyword map: trigger phrases resolve a plain-English question to a verb deterministically (the longest matching phrase wins, so "is it a buy" beats a bare "buy"). Only the top verbs are keyworded; anything else still resolves from an explicit verb token or the semantic routing table.

python3 scripts/router.py route "is NBIS a value trap?"   # -> risk
python3 scripts/router.py route "any red flags in PLTR?"  # -> redflags
python3 scripts/router.py route "how does AMD compare to NVDA"  # -> compare

If the question sounds like…	Trigger phrases (examples)	Routes to
is it cheap / a buy	is it a buy, overvalued, undervalued, fair value, should i buy, cheap	valuation
is it safe	is it safe, value trap, blow up, go bankrupt, too much debt, risky	risk
anything wrong	red flag(s), warning sign, going concern, anything to worry	redflags
can it survive	financial health, balance sheet, cash runway, solvency, self-funding	health
is it growing	growth rate, top line, is it growing, decelerating, accelerating	growth
does it have an edge	a moat, competitive advantage, pricing power, defensible	moat
what's it worth	intrinsic value, discounted cash flow, fair price, dcf, worth	dcf
is growth efficient	rule of 40, rule40, r40	rule40
which is better	compare, versus, vs, head-to-head, better than	compare
tell me everything	tell me about, walk me through, full picture, deep dive	company

route returns nothing when no phrase matches (then the agent falls back to the semantic routing table); a leading question word like what/how never hijacks routing. The map lives in KEYWORDS in scripts/router.py.

Slash commands

Plain English always works (Layer 1). But the verbs are the primary interface — a polished CLI over one shared engine, so numbers never diverge across commands.

Command	Answers	Backed by
/finance-skills <question>	anything — routes to the right lens	analyze + intent routing
/finance-skills company <ticker>	"tell me about this company" — a guided walkthrough	scripts/company.py
/finance-skills analyze <ticker>	"should I invest?" end-to-end (flagship)	scripts/analyze.py
/finance-skills framework <name> <ticker>	"run the SaaS / neocloud / semis lens"	scripts/framework.py
/finance-skills valuation <ticker>	"is it cheap?"	DCF + EV/EBITDA + Rule 40
/finance-skills dcf <ticker>	"what's it worth?"	DCF slice of analyze
/finance-skills rule40 <ticker>	"is growth efficient?"	segment-aware Rule of 40
/finance-skills growth <ticker>	"is it growing?"	growth + margins + regime
/finance-skills risk <ticker>	"what could go wrong?"	leverage, FCF, dilution, capex gap
/finance-skills redflags <ticker>	"any warning signs?"	scripts/redflags.py (severity-ranked flags)
/finance-skills health <ticker>	"can it survive?"	scripts/health.py (leverage, runway, dilution)
/finance-skills moat <ticker>	"does it have a durable edge?"	margins vs peers + references/
/finance-skills fiveforces <ticker>	"how good is the industry structure?"	Porter, applied to engine evidence
/finance-skills compare <a> <b> [...]	"which is better?"	scripts/compare.py (side-by-side table)
/finance-skills screen "<rule>" [tickers]	"which pass this filter?"	scripts/screen.py (field op value)
/finance-skills watchlist add|list|run <verb>	"track a set, run any verb across it"	scripts/watchlist.py
/finance-skills export <ticker> --format md|json|csv	"give me a shareable file"	scripts/export.py
/finance-skills learn <concept>	"what is a Magic Number / DCF / NRR?"	scripts/learn.py (offline)
/finance-skills help	grouped command help	scripts/router.py help

Shorthand and typos resolve instead of erroring (co→company, fw→framework, val→valuation, r40→rule40, 5forces→fiveforces, vluation→valuation).

company — example output (the guided walkthrough)

/finance-skills company CRWV steps through the business top-to-bottom, each stage flowing into the next, and ends with a synthesised verdict:

At a glance: the 9 stages for CRWV — strong margins (gross 70%) but BELOW its Rule-of-40 bar, negative FCF, and a verdict that hinges on backlog & funding runway.

═══ CoreWeave, Inc. (CRWV) — company walkthrough ═══
Source: fixture · as of 2026-Q1  [SAMPLE DATA — not live]
Price: $100   Market cap: $48.00B

■ Business Model
    Sector: Technology / Information Technology Services
    AI neocloud/hyperscaler — extreme growth funded by heavy GPU capex; judged on cash burn and backlog, not headline EBITDA.
        ▼
■ Competitive Advantage
    Gross margin 70.0% — high; suggests pricing power or a software-like cost structure.
    EBITDA margin 56.0% — operating leverage already showing.
        ▼
■ Revenue Drivers
    Revenue: $1.90B
    Growth (YoY): 111.1%.
        ▼
■ Margins
    Gross: 70.0%   EBITDA: 56.0%   FCF: -315.8%
    Negative free cash flow — growth is consuming cash (normal for the regime, watch runway).
        ▼
■ Financial Health
    Net debt: $11.50B
    Net debt / EBITDA: 10.81x — elevated; watch refinancing and covenants.
        ▼
■ Growth
    Growth rate: 111.1%
    AI neocloud/hyperscaler regime.
        ▼
■ Valuation
    DCF: DCF skipped: free cash flow is not positive (typical for capex-heavy growth names).
    EV/EBITDA: 55.9x.
        ▼
■ Risks
    Capital-intensity gap 372 pts — growth is capex-funded, not organically profitable.
    Below its Rule-of-40 bar (judged -668 vs 38).
    Share dilution 9.1% YoY — growth partly 'bought' with equity.
    Cash burn — depends on continued access to funding.
        ▼
■ Final Verdict
    A capital-intensive neocloud: the story is backlog and funding runway, not this quarter's margin.
    Falls short of its Rule-of-40 bar today.
    No DCF (FCF not positive), so lean on Rule-of-40 and multiples instead of intrinsic value.
    Not a recommendation — verify against primary filings before acting.

valuation — "is it cheap?" as a table

/finance-skills valuation <ticker> lays the valuation slice out as a scannable Metric | Value | Read table (DCF, EV/Sales, EV/EBITDA, Rule 40), flagging a distorted EV/EBITDA when EBITDA margin exceeds 100%:

═══ CoreWeave, Inc. (CRWV) — valuation ═══
Source: fixture · as of 2026-Q1  [SAMPLE DATA — not live]

  Metric              Value        Read
  ───────────────────────────────────────────────────────────────────
  Price               $100         —
  Market cap          $48.00B      —
  Enterprise value    $59.50B      market cap + net debt
  EV / Sales          31.3x        extreme — priced on growth, not sales
  EV / EBITDA         55.9x        expensive
  DCF / share         n/a          FCF negative — DCF skipped
  Rule of 40          -668 vs 38   BELOW BAR (ai neocloud)
  Revenue growth      111.1%       hypergrowth
  FCF margin          -315.8%      cash burn — depends on funding
  Net debt / EBITDA   10.81x       elevated — watch refinancing

Verdict: No DCF (FCF not positive), so it can't be anchored to intrinsic value — expensive on EV/Sales 31.3x; a growth/backlog bet, not supported by current cash flows.

framework — run a whole lens at once (honest about data)

/finance-skills framework saas CRWV runs every SaaS metric instead of making you pick. Metrics that need a disclosed KPI not in the financial statements (Magic Number, CAC payback, NRR) are flagged with their definition — never faked:

At a glance: computed from filings — Rule of 40 BELOW BAR, gross margin 70.0%, EV/EBITDA 55.9x; Magic Number, CAC payback & NRR flagged needs disclosed KPI rather than fabricated.

═══ CoreWeave, Inc. (CRWV) — SaaS / software quality framework ═══
Source: fixture · as of 2026-Q1  [SAMPLE DATA — not live]

  Metric                        Value / status
  ───────────────────────────────────────────────────────────────────
  Rule of 40                    judged -668 vs 38 bar → BELOW BAR (EBITDA 167 / FCF -205, gap 372)
  Gross margin                  70.0%
  FCF margin                    -315.8%
  Revenue growth (YoY)          111.1%
  EV/EBITDA                     55.9x
  Magic Number                  ⚠ needs disclosed KPI
  CAC payback                   ⚠ needs disclosed KPI
  Net revenue retention (NRR)   ⚠ needs disclosed KPI

  Not in the financial statements — check the 10-K / investor deck (defined, not faked):
    • Magic Number — net-new ARR ÷ prior-quarter S&M spend; >0.75 = efficient growth. Needs S&M + ARR disclosure.
    • CAC payback — months of gross-margin-adjusted revenue to recover customer acquisition cost. Needs S&M + new-customer/ARR disclosure.
    • Net revenue retention (NRR) — expansion − churn on existing customers; >120% is elite. A disclosed KPI, not in the financial statements.

Frameworks: saas, neocloud, semiconductor (python3 scripts/framework.py list).

learn — teach the concept, no ticker needed

/finance-skills learn dcf (also rule40, magic-number, nrr, five-forces, …):

═══ dcf ═══
Discounted cash flow: a company is worth the present value of its future free cash flow.

How to compute / read it:
  Two-stage model: grow FCF for N years, discount each year back, add a Gordon terminal value, subtract net debt, divide by shares. Here growth is a heuristic (trailing revenue growth, capped), discount 10%, terminal 3%.

Common trap:
  Output is only as good as the assumptions — tiny changes in growth/discount swing it wildly. Treat it as a rough anchor, and note it's skipped when FCF is negative.

analyze — example output

/finance-skills analyze CRWV (shown on the offline sample via --fixture; live output has the same shape with a yfinance source + timestamp):

At a glance: CRWV — AI-neocloud regime · Rule of 40 BELOW BAR (-668 vs 38) · burning cash (FCF -315.8%) · leverage 10.81x.

═══ CoreWeave, Inc. (CRWV) ═══
Source: fixture · as of 2026-Q1  [SAMPLE DATA — not live]
Sector: Technology / Information Technology Services
Price: $100   Market cap: $48.00B

Fundamentals (derived):
  Revenue growth (YoY): 111.1%
  EBITDA margin: 56.0%   FCF margin: -315.8%
  Capex intensity: 463.2%   Share dilution: 9.1%
  Net debt: $11.50B

Rule of 40 — regime: ai neocloud
  EBITDA-based: 167   FCF-based: -205   capital-intensity gap: 372
  Capex-adjusted: -668   dilution-adjusted: -677
  Judged on -668 vs benchmark 38 → BELOW BAR
  Verdict: Capital-intensive: growth is burning cash faster than it earns; watch backlog/RPO and funding runway.
    • Neocloud regime: the EBITDA-based score overstates health; judging on the capex-adjusted FCF score to reflect real GPU capital burn.
    • Large capital-intensity gap (372 pts) — growth is capex-funded, not organically profitable.

DCF: DCF skipped: free cash flow is not positive (typical for capex-heavy growth names).
Leverage: net debt / EBITDA = 10.81x

────────────────────────────────────────────────────────────
Read-only market analysis for research/education. Not investment advice; no trades are placed. Verify figures against primary filings before acting.

The valuation, growth, risk, and moat verbs run the same engine and lead with the matching slice of that report (e.g. risk leads with leverage 10.81×, the 372-pt capital-intensity gap, and dilution).

help — example output

/finance-skills help:

finance-skills — ask in plain English, or use a verb.

Top verbs:  company  analyze  valuation  framework  compare  learn

By question:
  Whole company          →  company, analyze, framework
  Is it cheap?           →  valuation, dcf, rule40, benchmark
  Is it safe?            →  risk, redflags, health
  Will it grow?          →  growth, opportunities, earnings
  Does it have an edge?  →  moat, fiveforces, management
  How does it compare?   →  compare, competitors, industry
  Learn a concept        →  learn
  Sector-specific        →  semiconductor, ai-cloud, banking, reit, insurance
  Power tools            →  screen, rank, portfolio, watchlist, export

Shorthand works too: val→valuation, r40→rule40, comp→compare, semis→semiconductor.
Typos are tolerated (e.g. 'vluation' → valuation).

Natural-language front door — example

/finance-skills Do you think NBIS and CRWV is a buy? first extracts the tickers, then runs the engine per ticker:

$ python3 scripts/router.py tickers "Do you think NBIS and CRWV is a buy?"
NBIS CRWV
$ python3 scripts/router.py r40
r40 → rule40 (alias)

Architecture

finance-skills/
├── SKILL.md                # skill entry: triggers, safety, invocation contract
├── install.sh              # install as an agent skill (claude/antigravity/codex)
├── pyproject.toml          # PEP 621 packaging (installs scripts/ as `finance_skills`)
├── scripts/                # sources — also the importable `finance_skills` package
│   ├── __init__.py         # package marker + __version__
│   ├── _entry.py           # console entry: `finance-skills` -> router.main(argv)
│   ├── data.py             # yfinance fetch + normalise + 6h cache + graceful fallback
│   ├── metrics.py          # PURE engine: segment-aware Rule 40, DCF, EV/EBITDA, Altman Z, Piotroski
│   ├── analyze.py          # orchestrator: fetch → compute → report (flagship `analyze`)
│   ├── company.py          # 9-stage sequential walkthrough (view over analyze)
│   ├── framework.py        # named frameworks as checklists (saas/neocloud/semiconductor)
│   ├── redflags.py         # warning-sign scan with severity (view over analyze)
│   ├── health.py           # solvency: leverage, cash runway, dilution (view over analyze)
│   ├── compare.py          # side-by-side table for two+ tickers
│   ├── screen.py           # filter a set of tickers by a tiny `field op value` rule
│   ├── watchlist.py        # saved named ticker lists; run any verb across them
│   ├── export.py           # render a verb to a Markdown / JSON / CSV file
│   ├── learn.py            # offline concept explainers (no ticker, no network)
│   └── router.py           # ticker extraction + alias/fuzzy + keyword→verb routing (pure)
├── references/
│   ├── rule40.md           # segment-aware Rule of 40 methodology + benchmarks
│   └── ai-cloud.md         # AI-cloud/neocloud sector framework (capex, backlog/RPO)
├── tests/                  # offline unit tests (pure math + orchestrator + router + CLI)
└── requirements.txt        # yfinance

The same files serve two roles. Run in place, python3 scripts/<mod>.py drives the agent skill (what SKILL.md / install.sh use). Installed from PyPI, the directory is remapped to the importable finance_skills package with a finance-skills console command — the intra-module imports resolve both ways via a small try: from finance_skills import … / except ImportError: import … shim.

The engine is one source of truth: metrics.py is pure and offline-testable; data.py is the only module that touches the network; analyze.py composes them. Every specialised command (company, framework, valuation, dcf, rule40, risk…) is a view over analyze, so numbers never diverge between commands.

Install as a Python package

pip install finance-skills          # console command + importable engine
finance-skills help                 # grouped command help
python -c "import finance_skills"   # the engine as a library

CLI usage (also drives the skill)

pip install -r requirements.txt

python3 scripts/analyze.py NVDA            # full live report
python3 scripts/company.py NVDA             # guided 9-stage walkthrough
python3 scripts/framework.py saas NVDA      # run the SaaS lens as a checklist
python3 scripts/learn.py rule40             # explain a concept (offline)
python3 scripts/analyze.py CRWV --fixture   # offline sample (no network)
python3 scripts/router.py tickers "is NBIS a buy?"   # -> NBIS
python3 scripts/router.py help              # grouped help

Platform note

Live fetching uses yfinance (network) → works on Claude Code and locally, not on the Claude.ai sandbox. Without network, use --fixture (CRWV, NBIS samples, clearly labelled non-live) or the skill will say live data is unavailable.

Development

python3 -m pytest tests/ -q     # 118 offline tests (no network needed)

• tests/test_metrics.py — regime classification, dual-margin/capex-adjusted Rule 40 (locks the CoreWeave/Nebius examples), DCF guards, Altman Z, Piotroski.
• tests/test_analyze.py — orchestrator on fixtures + graceful no-data path.
• tests/test_data.py — statement column-ordering + net-debt fail-closed behaviour.
• tests/test_company.py — the 9 walkthrough stages, in order, with data-gap flags.
• tests/test_framework.py — computed metrics vs honestly-flagged disclosed KPIs.
• tests/test_learn.py — concept/alias/fuzzy resolution for the explainers.
• tests/test_router.py — ticker extraction, alias/fuzzy + keyword→verb routing, grouped help.
• tests/test_redflags.py / test_health.py / test_compare.py — the new engine views.
• tests/test_screen.py — the safe field op value rule parser and fail-closed missing data.
• tests/test_watchlist.py / test_export.py — persistence and md/json/csv output.
• tests/test_entry.py — the finance-skills console entry point runs help.

Status & roadmap

The real engine proven end-to-end on live data plus offline fixtures, installable as a cross-tool skill, with a verb-first CLI (company, framework, learn, …) layered over it. Next, over the same engine: more sector references (semiconductor.md, banking.md, reit.md), screen/rank views, trend arrows, and backlog/RPO ingestion to light up the framework KPI rows.