heurist-finance: Skill for Claude Code

Trade Postmortem

Shared plumbing and output contract: ../_shared/analyst-core.md. The feedback loop is where a trader either compounds skill or compounds superstition — and P&L alone teaches the wrong lessons, because markets pay bad process often enough to reinforce it. This skill closes the loop that trading-plan opens: the plan documents in trading/plans/ are the exam answer key; the postmortem grades against them.

The 2×2 (grade every trade on both axes)

The paid-mistake quadrant is the reason this skill exists: profitable rule-breaking is the most reinforcing event in trading and the most corrosive. Say it plainly when found.

Hindsight-proof attribution (the hard part done honestly)

Judging a decision by information available after it is the universal postmortem failure. Reconstruct the as-of-then view before grading:

• What was knowable at entry: the plan document's thesis and scenario table; the as-of-then narrative, reconstructed from dated articles via WebSearch; macro as then reported — use fred macro_vintage_history (this is exactly what vintage data exists for: the payroll number the user traded on, not its later revision); prices/IV at the time from prices.
• If the reported fills conflict with the verified tape (a price that never traded that day, a date with no such move): don't silently accept or correct. Grade the reported structure on process, flag the tape conflict explicitly, and cross-check whether the narrative's shape (direction, rough timing) survives on the real bars — if it doesn't, the postmortem's first finding is the record-keeping.
• Then the four-way attribution, each scored right/wrong given that reconstruction: Thesis (did the predicted mechanism operate?), Timing (right thesis, wrong clock is the most common split verdict — note it feeds sizing, not thesis, next time), Sizing (did size force a behavioral exit before the thesis resolved? cross-ref position-sizing's 11pm test), Exit (against the plan's invalidation list: did they exit on signal, on noise, or on pulse? — the field card's noise/signal split, applied retroactively).
• Counterfactual honestly bounded: "would've made X if held" uses the plan's exit rule as the counterfactual, never the peak price — the peak was only visible in hindsight, and grading against it manufactures regret, which drives the next oversize.