8 sessions on Git with a coding agent: the agent failed at least once in 2, and 4 finished without a human stepping in.
Takeaway: Next time I'd just tell it upfront that this is a Claude Code skill repo (not a plugin marketplace) and that it needs SKILL.md copied into ~/.claude/skills/, since the agent burned several turns discovering that itself.
Takeaway: Next time I'd tell it up front to match the existing convention in develop before inventing new schema entries like licenses, since checking real precedent is what actually fixed things.
Takeaway: I'd start the next audit by explicitly telling it to re-verify the NEVER DO list against reality before touching anything, since two of the entries there turned out to be stale (pacnew notes, BIOS assumption).
Takeaway: I'd start by pointing straight at the dotfiles repo and naming the waybar/walker/mako/hypr files instead of asking it to go find my 'SYSTEM design' — that search cost real time.
Takeaway: Next time I'd just say up front that it's a skill repo, not a plugin marketplace, since the agent burned a few steps trying the marketplace/plugin install path before checking the repo contents directly.
Takeaway: Next time I'd tell it upfront to match the existing sibling model entries' conventions (license, fyi text) instead of inventing new fields, since that's what caused the rework.
Takeaway: I'd start the next prompt by telling it to always check the NEVER-DO list for staleness before treating any entry as gospel — that's what let us safely revisit both the mkinitcpio hook and the 'never update BIOS' rule instead of leaving good gains on the table.
Takeaway: Next time I'd start the audit prompt with 'compare live config to the dotfiles repo and flag drift' up front, since half the useful findings came from that diff rather than the visual review.