Fourth in the series is align, a small single-plugin marketplace that treats “the model said something wrong” as a problem worth building tooling around, rather than just something you fix in the moment and move on from.
Three commands
The skill centres on three slash commands. /align rates each claim Claude made against a six-way canonical scale – correct, wrong, almost, needs nuance, can’t verify, or skipped – using a local HTML form, and archives any corrections as structured Markdown rather than letting them evaporate at the end of the chat.
/diagnose is the more interesting one: it traces a wrong claim back to the stale instruction or outdated context that likely caused it, using quoted-evidence citations, a 0–100 confidence score, and a threshold-80 filter to avoid flagging things it isn’t sure about. It’s read-only, so it doesn’t touch your files.
/retro takes this a step further by clustering patterns across multiple past sessions and opening patch pull requests to fix the underlying instructions – gated behind a human review at every patch, not auto-merged.
Why this is a different shape of tool
Most correction workflows are one-off: you tell Claude it’s wrong, it apologises, you move on. Align is aimed at the same wrong answer not happening again next week because the CLAUDE.md or skill instruction that caused it never got fixed.
Installing it
It’s MIT licensed and runs entirely locally with no external service:
/plugin marketplace add ggrigo/alignNext Sunday: the andrej-karpathy-skills repo – a single CLAUDE.md file that picked up over 100,000 GitHub stars for encoding four behavioural rules distilled from Andrej Karpathy’s own complaints about coding agents.
Leave a Reply
You must be logged in to post a comment.