compound-learnings

name	compound-learnings
description	The closing step of every task — codify what was learned so the system gets better each use. Use after completing any task where the agent was corrected, a bug was found, a procedure was discovered, or friction repeated. Also on request ("capture this", "make sure we don't repeat that").
source	Kieran Klaassen (compound engineering — the fourth step); Mitchell Hashimoto (mistake → guardrail reflex); Boris Cherny (CLAUDE.md as error-ledger)
am_pattern	A1, C1
triggers	["lessons learned","compound learnings","postmortem","post-mortem","capture what we learned"]

"The first three steps produce a feature; the fourth produces a system that builds features better." You are running the compound step: turn this task's lessons into durable artifacts so no one — human or agent — pays for the same lesson twice.

Procedure

1. Harvest. Scan the task you just finished for compoundable material:

   • a mistake the agent made (or almost made) that a rule could have prevented
   • a correction the human gave more than once
   • a procedure you had to figure out (how to run X, how to test Y, where Z lives)
   • a command sequence you ran more than twice
   • a bug class the checks didn't catch