What I found
A first-draft equity research report is repeatable work: pull the SEC filings, compute the ratios, build a DCF, compare multiples, survey the material ESG disclosures, and write a thesis. The framework is standard and the data is public. So I asked a narrower question: could an agent produce a defensible first draft, on free primary sources, for a couple of dollars? It could, but only if I encoded the judgment, not just the steps.
What I built
A reusable skill, company-performance-analysis, that runs locally on the OpenClaw agent framework with Anthropic's Claude. It is a folder of markdown, about 44 KB in all: one trigger-and-workflow file and four reference files that carry the analytical method.
company-performance-analysis/
SKILL.md trigger, scope check, 7-step workflow
financial-ratios.md formulas, 3-year trend lens, sector benchmarks
valuation-frameworks.md WACC, 5-year DCF, reverse DCF, multiples, sector pivots
esg-checklist.md materiality lens, 10-K extraction, governance red flags
report-template.md the 6-section report and citation discipline
It reads SEC EDGAR, Yahoo Finance, StockAnalysis.com, and company filings, and returns an 1,800 to 2,500 word report: company snapshot, a 3-year financial health table, a valuation with a DCF and a reverse DCF and a sensitivity grid, a peer comparison, material ESG, and a Buy, Hold, or Sell with a 12-month target. Every number is cited to a primary source.
How it works
The same skill, run across five sectors. Pick one to see how it adapted, and open the Tesla run to see the reverse-DCF decomposition from the live demo.
The same skill, run on 4 sectors. Pick one to see how it adapted.
Method: DCF + reverse DCF
The live demo. An audience member picked TSLA, unseen. In about five minutes the skill produced a 2,400-word report with a Sell and a $200 target, breaking a three-Hold streak through its anti-confirmation-bias guardrails.
Reverse-DCF decomposition
A DCF of the current auto and energy businesses came to about $30 a share. The reverse DCF at the market price implied operationally impossible assumptions, roughly $3.8T of Year-5 revenue, showing that about $1.5T of market cap was pure optionality on pre-commercial robotaxi and humanoid-robot businesses.
Recommendations across the tests: Hold, Hold, Hold, Sell. Calibrated to resist the default-to-Buy bias, on OpenClaw with Anthropic's Claude.
The bank test is the one I am proudest of. The skill refused to run a DCF on JPMorgan, because financing is the product and cash-flow definitions break down, and pivoted to a Justified P/TBV with a dividend-discount cross-check, swapped debt-to-equity for the CET1 capital ratio, and chose a bank peer set. It announced the pivot up front. That is methodological judgment encoded in a reference file, not template substitution.
Impact
- Validated across five sectors with recommendations that distributed, Hold, Hold, Hold, then a Sell on Tesla, rather than defaulting to Buy.
- About $2 and 5 minutes per report, and under $10 for the whole project after I set a hard spend cap.
- I built and tested it on the top-tier model, then moved production to a faster, lower-cost tier after side-by-side testing showed equivalent quality for roughly a fifth of the cost. Because the model is decoupled from the skill, that was a one-line config change.
The transferable lesson
Encode the judgment, not just the steps. The parts that made this more than generic AI output were the guardrails: a calibration note so it does not default to Buy, sector-aware methodology, a required reverse DCF as a check on its own narrative, and limitations stated plainly rather than buried. And one lesson I did not expect: operational guardrails matter as much as analytical ones. A background process quietly ran up about $80 in unintended API charges before I caught it, diagnosed the cause, and set a hard cap. The deliverable was never the reports. It was the skill, and the method it carries.