Skip to main content

​
Summary

Built Grace’s self-improvement CI/CD layer by adapting two proven patterns: the ClawHub pskoett/self-improving-agent skill (.learnings/ directory + hooks) and the existing 5-repo archive system. The result is a 4-tier promotion pipeline where errors are captured automatically, learnings accumulate over sessions, and mature patterns get promoted to permanent agent behaviour.

​
What changed operationally

Three Claude Code hooks now fire automatically on every Grace session: InstructionsLoaded injects recent learnings as context, PostToolUseFailure captures Bash errors to .learnings/ERRORS.md, and SessionEnd logs session summaries to .learnings/LEARNINGS.md. Grace is now registered in the archive system via .archive-manifest.yaml, enabling learnings to route to atlas and findings to so1-content. The PROMOTE trigger word lets AJ explicitly control when learnings get promoted from raw capture to permanent agent files — respecting the ADHD-friendly gating workflow rather than running autonomous cron jobs.

​
Business impact

  • Grace now gets smarter with every session — errors accumulate, patterns emerge, and AJ decides which ones become permanent behaviour
  • The archive system grows from 5 repos to 6 (grace now participates)
  • Self-improvement is auditable — .learnings/ is git-tracked, promotions are documented
  • Zero-effort capture — hooks fire without AJ doing anything different

​
Operational takeaway

The ClawHub self-improving-agent pattern works, but needs two adaptations for a solo founder with ADHD: (1) promotion must be human-gated (not cron) to prevent scope anxiety, and (2) the capture layer must be completely silent — hooks that require attention on every session get disabled within a week.