Type R · own timeConviction

Compiling agent experience into a persistent skill wiki

proposed
voting
running · 87d
measuring
concluded

Preregistration · v1 · 8 Jun 2026 · immutable after start

An extension creates a new version rather than editing the old.

Hypothesis

An agent that writes what it learned from each coding session into a human-readable skill page, and reads the relevant pages at the start of the next, repeats fewer known mistakes than one that starts cold.

Kill condition

Type R: none required. Logged at the end, one sentence minimum. Standing cap $200/month, no client data.

MethodA day a fortnight over a quarter. Claude Code hooks that write a skill page per session on the lab's own repos and load the top matching pages on start. Count repeated mistakes by hand from session logs, before and after.

Expected costUnder $200/month, one day a fortnight

Expected durationOne quarter

Running notes · fed from harness sessions and by the pair

harness31 Aug 2026AW Adam Witanowski
58 pages after consolidation; hand count of repeated mistakes on 3 repos: 14 in the fortnight before, 6 in the last
manual3 Aug 2026AW Adam Witanowski
Pages are getting long and contradicting each other. Added a weekly consolidation pass; it is the same problem as the episodic store, one level up.
harness6 Jul 2026AW Adam Witanowski
hooks: post-session writer + pre-session loader; 31 skill pages; loader picks by path + recent edits
manual8 Jun 2026AW Adam Witanowski
Conviction run, my name on it, no evidence claimed. I think consolidation with a human-readable output is the piece the memory field is missing.

Harness notes are auto-captured from Claude Code sessions: model, date, commit, session reference. Never the transcript, code or paths.

Artifacts

reposkill-wiki hooks for Claude Coderepo

noteHand-counted repeated mistakes by fortnightnotebook