Mine session corrections before proposing new Claude rules
At a glance
Artem Zhutov used Claude Code to turn recurring corrections from past sessions into an HTML report of proposed instruction changes. This workflow is listed under agent-memory, instruction-management, workflow-review.
Zhutov asked Claude to examine his last 50 sessions. Ten extraction agents divided the history, another stage grouped recurring corrections, and final agents reconciled proposed rules against his existing CLAUDE.md. He requested an HTML report so he could inspect the corrections alongside the proposed edits.
How it went
He reports that the run analyzed 49 sessions and extracted 86 corrections. The report paired proposed changes with his corrective comments and evidence from the underlying sessions.
Worth knowing
After inspecting the proposals, he preferred packaging recurring fixes into skills rather than putting every finding into CLAUDE.md.
Save ideas in this browser. Upvotes count across the board.
Open-source local procedural memory that learns only after tests pass, then hands Claude Code/Codex/Cursor a short verified skill note on similar tasks (reported token/tool cuts in published benchmarks).
A Technically writer documents a month-long Claude Code experiment managing a Meta Ads account: fresh sessions read prior logs, pull performance data, decide, act or abstain, and commit the updated history.
A founder describes a working beta that scans a repository, asks product questions, and writes a behavior contract distinguishing confirmed decisions from unsettled assumptions.
An indie hacker packaged an agency's eight repeated workflows (briefs, client approval with revision counters, quotes with margin checks, content calendars, post-mortems) as free MIT Claude skills.
