Menu
Score a bundled model's answers against the gold finance QA set and report honest metrics + calibration. Usage /evaluate [predictions.json]
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Value a public company's equity end-to-end (FCFE at CAPM cost of equity) with an IB deal team — multi-year EDGAR data, scenario forecasting, Monte-Carlo DCF, dual-source comps, a verification gate, and an investment-committee sign-off. Run `/valuation <cik|ticker>`.
Value a public company end-to-end from its CIK or ticker. Run `/valuation <cik|ticker>`. Orchestrates EDGAR fetch, three parallel valuation lanes (Monte-Carlo DCF, dual-source comparables, qualitative/risk), and reconciliation into a fair-value median + P10–P90, benchmarked against the real market price.
Generate a professional 2-page report on a publicly-traded company as a self-contained HTML website. Given a company name, orchestrates five subagents in order (Generalist, Data Retriever, Data Scientist, Financial Analyst, Report Writer) — pulling the latest 10-K from SEC EDGAR, cleaning the financials, analyzing them, and writing the report. Use when the user asks for a company report, equity write-up, or 10-K-based analysis. Triggers include "write a report on <company>", "company report for <X>", "/company-report".
One sentence on what this command does and when to invoke it. Usage /my-skill "<arg>"
Answer a question about the bundled NovaCorp filings, grounded and graded. Usage /ask "<question>"
Route a research question, gather grounded evidence, and write a cited brief. Usage /brief "<question>"
| name | evaluate |
| description | Score a bundled model's answers against the gold finance QA set and report honest metrics + calibration. Usage /evaluate [predictions.json] |
Run the full evaluation and write a grounded report. Default predictions file:
data/candidates_overconfident.json. The tools own every number; you only run them,
read them back, and explain them.
Score (scorer agent):
python -m tools.score --gold data/gold.json --pred <predictions.json> --out reports/_scored.json
Record accuracy, mean_f1, n, n_correct.
Calibrate (calibration-analyst agent):
python -m tools.calibration --scored reports/_scored.json
Record ece, mean_confidence, the confident_wrong list, and the reliability bins.
Report (reviewer agent): write reports/<model>.md with, every figure cited from
the tools above:
Rules: report the numbers honestly; never round away a failure. Do not call an
overconfident model "well calibrated" because its accuracy looks acceptable. If the
confident_wrong list is non-empty, it leads the report.
Try these to start:
/evaluate → evaluates data/candidates_overconfident.json; expect ~0.5 accuracy, high ECE, five confident-wrong cases./evaluate data/candidates_calibrated.json → a model whose confidence tracks its accuracy; expect low ECE and no confident-wrong cases.