latour-ai·latour-ai/skills/claim-sentiment-extractor/SKILL.md

claim-sentiment-extractor

Skill69 starsChanged 25 days agoNo licenceView on GitHub

Turn any research document — an annual, semi-annual, or quarterly report, an earnings-call transcript, a broker note, an expert-network transcript, an investor-day deck — into a claim-level table. Each row is one claim with the verbatim quote and its location, a fact-or-sentiment label, a polarity (positive, negative, neutral), a theme, and a score against the named thesis factors for the position, such as expected buybacks, margin expansion, or deleveraging. The labelling is calibrated to the analyst's own judgment: the analyst keeps a plain file of labelled example claims and corrections, the skill reads it on every run, and every run reports how far its labels agree with the analyst's. For research analysts and portfolio managers on long/short, long-only, activist, and credit teams. Trigger on phrases like "pull the claims out of this transcript," "fact versus sentiment on this note," "score this report against our thesis factors," "tag every claim in the semi-annual," "how bullish is this broker really, claim by claim," "run the claims table and show agreement with my labels," or when a user attaches a document plus a list of thesis factors. Do NOT use to track how management's language shifts across quarters. Do NOT use to file whole items in the incoming flow as bull or bear evidence on a position's debates — use the key-debates-ledger skill. Do NOT use for a structured note on a single expert call with a credibility assessment. Do NOT use to decide which sell-side notes deserve reading — use the sellside-research-triage skill. Do NOT use for what the market has priced in ahead of a print. Quotes are verbatim, facts and sentiment are never blended in one label, and the table carries no trade recommendation.