finance-research-agent: Skill for Claude Code

Screen to shortlist

Turn the universe into a ranked, study-ready shortlist the user can act on. Producing the shortlist IS the deliverable — do not refuse it. The user does further study and makes the final call; you provide the evidence and the ranking.

Consult the investing-principles skill for the ready-made strategy presets (Coffee Can, Magic Formula, Graham Defensive, QARP), the named quantitative scores (Piotroski/Altman/Sloan/ DuPont/Graham Number), the peer-comp quality discipline, and how to weigh quality vs value — use one of those as the screen's backbone unless the user specifies their own criteria.

The verdict-taxonomy and screen-integrity practices below are adapted from Anthropic's public anthropics/financial-services reference repo (Apache 2.0) — its private-equity/deal-screening and equity-research/idea-generation skills. Core line from that repo worth repeating here: "screens surface candidates, not conclusions — every screen output still needs fundamental work." A shortlist is where study starts, not where it ends.

Procedure

1. Clarify criteria, then commit. If the user's criteria are vague, propose concrete thresholds (valuation, quality, growth, momentum) — or offer a named strategy preset from the investing-principles skill — state them plainly, and proceed. Ask only if a genuinely blocking choice remains.

2. Ethical exclusions FIRST — screen_stocks(exclude_categories=[...]), using ONLY categories the user has explicitly confirmed. Never assume a category they haven't named. Relay the exclusion counts and reasons so it's transparent what was dropped and why.

3. Quantitative filter in the SAME call — pass the user's numeric criteria to the same screen_stocks call; exclusions and filters compose in one pass. For a sector-scoped screen, use the sector/industry argument (or sector_analysis first to size the field).

3b. A specific past YEAR, not today's snapshot ("ROCE > 20% in FY2024", "best/worst performers in 2023") → screen_by_year(year, kind="fundamental"|"price_return", min/max, ...) instead of screen_stocks — the latter only ever sees the latest values. For qualitative colour on any matched name for that exact period, follow up with search_documents(symbol=X, date_from=<FY start>, date_to=<FY end>, doc_types=["concall_transcript"]) scoped to that year/quarter — never an unscoped search when you already know the period.