nudge: Skill for Claude Code

Nudge Learnings

Purpose

Nudge learnings are repo-local debugging memory, not broad agent memory. They live in Git with the repo, so each branch or worktree only sees the notes present in that checkout. Use this skill proactively when a repo-specific failure might already have a recorded fix, and after solving a durable repo-specific issue that future agents should not rediscover.

When to use

• You are debugging a failing command, test, build, integration, dependency, local environment, or toolchain behavior in a repo with .nudge/learned.
• Nudge surfaces learned repo context.
• The user asks to search, list, record, or update Nudge learnings.
• You fixed a repo-specific issue that another agent could plausibly hit again.

Do not record generic programming advice, broad personal preferences, secrets, credentials, or one-off observations that are unlikely to recur.

Workflow

1. If .nudge/learned exists, search before deep investigation:
   • Use concrete symptoms, command names, package names, paths, tools, and copied error fragments.
   • Prefer several focused searches over one vague query.
   • If no learning matches, continue normal debugging.
2. If Nudge already surfaced a note, read the cited .nudge/learned/*.md file.
3. Apply a note only when its symptoms, environment, and fix match the current task. If it does not match, say why and keep investigating.
4. After fixing a durable repo-specific issue, record a concise note with nudge learn add and verify it with nudge learn search or nudge learn list.

Commands

nudge learn search "exact error text package command"
nudge learn list
nudge learn add --title "Short specific title" --body "What went wrong..."
nudge learn embeddings status

Note Shape

# Short specific title

## What went wrong

Name the command, tool, package, path, error, and symptoms.

## Fix

Give the exact fix and caveats.

## Verification

State the command or observation that proved the fix worked.