AugmentClaude

Correct

Fix repeated agent mistakes by restructuring code, adding checks, tests, and rules.

Installation

  1. Make sure Claude is on your device and in your terminal.

    Skills load from ~/.claude/skills/ when Claude Code starts up — so you need it on your machine first. If you don't have it yet, install it once with the command below, then run claude in any terminal to verify.

    One-time setup
    npm i -g @anthropic-ai/claude-code

    Already have it? Skip ahead.

  2. Paste into Claude Code or into your terminal.

    This copies the whole skill folder into ~/.claude/skills/correct-savannahfeder/ — the SKILL.md plus any scripts, reference docs, or templates the skill ships with. Works for every skill.

    What this command does
    1. Makes a new, empty temporary folder.
    2. Downloads the skill's folder, .claude/skills/correct, from github.com/savannahfeder/agentbox (branch main), the GitHub repository it's listed from. You get the version that is there now, not a copy we keep.
    3. Copies that folder to ~/.claude/skills/correct-savannahfeder/.
    4. Deletes the temporary folder.

    Installing runs none of the skill's files; it only downloads and copies them. When you later use the skill, Claude follows its SKILL.md and may run scripts that come with it. Read them first.

    How we check skills, and what we don't · Report this skill

    Faster alternative (instruction-only skills)

    Skips the clone and grabs only the SKILL.md file. Don't use this if the skill ships Python scripts, reference markdowns, or asset templates — they won't be downloaded and the skill will fail when it tries to load them.

    Quick install (SKILL.md only)
    Sign up to copy
  3. Restart Claude Code.

    Quit and reopen Claude Code (or any other agent that loads from ~/.claude/skills/). New skills are picked up on startup.

  4. Just ask Claude.

    Skills auto-activate when your request matches the skill's description — no slash command needed. Trigger phrases live in the skill's own frontmatter; you can read them in the “What this skill does” section above.

Prefer to read the source first? Open on GitHub.

When Claude uses it

Find the mistakes agents keep repeating in this repository and make each one impossible, by changing the code's structure first, then a check that fails loudly, then a test, and a written rule only last. Proves each new check goes red on a real past mistake. Use when asked to run /correct, or after the maintainer corrects agents for the same thing twice.

What this skill does

Correct

The maintainer keeps correcting agents in this repository for the same mistakes. Change the repository so the next agent cannot make them.

Assume every contributor is an agent that sees only the files it opened, copies the nearest example, and takes the shortest path that passes the tests. Make a change that looks right from one file be right for the whole repository.

This repository already works this way in places. tests/no-worker-drives-her-own-browser.test.mts turned a rule into a failing test, and scripts/hooks/ turned "never commit in the app's folder" into a refusal. This skill does that on purpose, across the history.

Find the mistake classes

Read, newest first:

  • git log for reverts, "fix again", "put back", and fixes to a fix.
  • CLAUDE.md, and the maintainer's private notes if CLAUDE.md points to them. Every section that tells an incident story is a class that was fixed with prose.
  • The maintainer's replies on threads in the Agentbox store (list_work_items, read_chat): look for "again", "I already said", "why did".
  • Review comments on pull requests.

Group what you find into classes. A class counts once it has happened twice.

Fix each class at the highest level that works

  1. Remove it with structure. One owner for each piece of state, one supported way to do each task, one copy of each rule (as shared/rank.mts is for ranking). Delete the old way and any dead code an agent would copy.
  2. Make the wrong thing fail loudly. A type, a runtime check, a git hook, or a check in scripts/ whose error names the file or function to use instead. If the pattern is already common, fail only when a change adds more of it.
  3. A test of the behaviour, named as a sentence and opened with what broke, as CLAUDE.md requires. Fix or delete any test that still passes when the behaviour breaks.
  4. A written rule, last, and only for a judgment call. Nothing fails when an agent skips a paragraph.

Prove it

Fix the most frequent classes first, one commit each. Prove each new check fails on a real past mistake: check out the file from the commit that had the mistake, or put the mistake back by hand, and watch the check go red. Paste that output. Run npm run test:changed before you report.

Nothing personal goes into a test or a check: no names, account ids, home paths or real messages. The repository is public.

Keep the rules and their checks together

Where CLAUDE.md states a rule that is now enforced, add one line naming the test or check that enforces it, the way the browser rule does. When a rule's mistake can no longer happen, shorten the prose to that pointer. When the maintainer corrects an agent for something CLAUDE.md already says and nothing enforces, that is a repeat: enforce it in the same change.

Reply

Each class, with its evidence (commits, threads, dates), the level you chose, why a higher level did not work, and the red-then-green proof.

Adapted from pstack's correct skill, https://github.com/cursor/plugins/tree/main/pstack (MIT).

Related skills