finance-research-agent: Skill for Claude Code

Investing principles & best practices

This is the "how a disciplined investor thinks" layer for Indian equities. Apply it on top of the findata tools: the tools give the raw evidence, this gives the judgement AND the exact formulas for turning that evidence into falsifiable scores. Distilled from The Intelligent Investor (Graham), The Little Book That Beats the Market (Greenblatt), The Little Book of Valuation (Damodaran), Coffee Can Investing (Mukherjea/Ambit — India-specific), The Four Pillars of Investing (Bernstein), The Behavioral Investor (Crosby), A Random Walk Down Wall Street (Malkiel), plus the academic accounting-quality/fraud-detection literature (Sloan 1996, Piotroski 2000, Altman 1968/2017 EM variant) that every serious forensic screen in the industry is actually built on. These are frameworks, not guarantees — markets are largely efficient, so edge comes from discipline and from catching what others don't check, not from certainty.

Some methodology below (peer-comp quality discipline, DD severity taxonomy, screening-verdict framing) is adapted from Anthropic's own public reference implementation for financial-services agents, anthropics/financial-services (Apache 2.0) — its equity-research, private-equity and financial-analysis skill verticals cover the same ground as this project at institutional-desk scale; adapted here for the local Indian-equities data lake rather than the Excel/DOCX-workbook pipeline that repo builds toward.

Ground rule for everything below: every score here is a computed composite of tool-sourced inputs — never a number pulled from memory, and never a number with a silently-assumed input. If screen_stocks/valuation_summary already expose a column directly, use it; where a score needs a ratio the tools don't pre-compute (Altman/Piotroski/Sloan), pull the raw lines from financial_statements (profit_loss/balance_sheet/cash_flow, 2+ years) and show the inputs plugged into the formula, not just the final number. State which figures came from which tool call. If a formula needs an input this data lake genuinely doesn't expose (no placeholder, no "assume unchanged"), say the score can't be computed and why — don't publish a number built on an assumed component.