20 Claude Code skills for falsification-gated self-improvement, grounded in FGSE paper

blaizedsouza · x · 2026-11-17

A thread pairs 20 practical Claude Code skills with an arXiv paper on falsification-gated agent memory. The core loop: lessons it keeps → checks it runs → edits it can trust.

Key skills include /lesson-as-hypothesis (save lessons with a test), /refine-after-episode, /falsifiable-rules, /write-own-verifier (write the checker first), /inspect-dont-trust, /replay-check (new code must replay old steps), and /holdout-gate (keep only what wins on held-out data).

The underlying paper, "When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents" (arXiv:2611.14326), proposes FGSE: represent each candidate lesson as a structured hypothesis with precondition, behavior change, expected effect and verifier; trial it in temporary state and admit it into persistent memory only after passing target-transfer, falsification and archived-regression tests — addressing the problem that local success is an unreliable signal for reusable, safe knowledge.

Original post →

More from coding & agent

coding & agent channel →