mcpField
C
correct

correct

github.com/yeachan-heoその他40.7k7 回インストール今日更新
リポジトリ

correct について

このリポジトリで Agent が繰り返す誤りを見つけ、最も強制力の高い層でそれぞれを防止します。lint ルール、型ガード、ランタイムチェック、スクリプト、最後にテキストの順で検討し、すべての防止策が実際に発動することを証明します。

インストール

コマンドをコピーしてターミナルで実行します。

$npx skills add https://github.com/yeachan-heo/oh-my-claudecode --skill correct

インストール後、Claude Code で/correctを入力して読み込まれたか確認します。

README（原文）English

プロジェクトの README より

この README の日本語訳はまだありません。以下は英語の原文です。日本語の紹介はページ上部をご覧ください。

Correct

Text instructions decay; structure reruns. This skill converts a repeated agent mistake into an enforcement the mistake cannot survive.

When to use

  • the same class of error appears twice or more (wrong import path, missed codegen step, broken invariant, slop pattern)
  • retrospective findings in .omc/notepads/ or review history name recurring friction

Workflow

  1. Collect. Mine the evidence — resolve OMC's state root first (the checkout's .omc/ by default; OMC_STATE_DIR or a parent .omc-workspace moves it, and a package-subdirectory launch must not read a stray checkout-local .omc): the notepads under <state-root>/notepads/*/problems.md, the verification ledger at <state-root>/state/verification-ledger.tsv, review comments, recent fix commits. Group by root cause, not by symptom.
  2. Rank. Frequency × cost. Take the top mistakes; one at a time.
  3. Pick the highest enforcement level that works:
    1. linter rule / compiler error — the mistake stops compiling
    2. type-level guard — the mistake stops type-checking
    3. runtime check or assertion — the mistake fails loudly at the boundary
    4. script or codemod — the correct action is one command
    5. project skill text — the last resort; note why every higher level failed
  4. Implement the enforcement with its exception path documented (how a legitimate case opts out).
  5. Prove it fires. Reproduce the mistake deliberately and show the enforcement catching it, with real output. An enforcement you have not seen fire is a hypothesis (an external benchmark-checklist discipline applies).
  6. Record honestly: the ledger verdict matches what the repro actually exercised — a lint rule or type guard firing is enforcement-verified, a runtime check exercised by a test is unit-test-verified, a proof is formally-verified. Never record a stronger verdict than the evidence.

Rules

  • Never weaken an enforcement to make a test pass; the enforcement is the requirement now.
  • Record what was encoded in the state-root ledger (<state-root>/state/verification-ledger.tsv per the collection step) with the verdict the repro actually exercised (see workflow step 6) and the repro command as evidence.
  • Text added at level 5 references this skill, so the next occurrence tries to climb the ladder again.

コメント

Markdown 対応