mcpField
F
figure-it-out

figure-it-out

github.com/yeachan-heoその他40.7k8 回インストール今日更新
リポジトリ

figure-it-out について

Figure It Out は、反復作業や高リスク作業向けの再利用可能なプレイブックを設計・検証・記録する OMC スキルです。確認可能な手順や失敗時の分岐を定め、実際のタスクで試行します。

主な機能

  1. 01実行可能な目標と確認手順
  2. 02各手順の失敗時対応
  3. 03実際のタスクでの試行
  4. 04プレイブックの検証と証拠記録

活用シーン

  1. 01反復するリリース手順の記録
  2. 02コンプライアンス調査の設計
  3. 03高リスクなデータバックフィルの計画

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 figure-it-out

インストール後、Claude Code で/figure-it-out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。

Figure It Out

No bundled skill fits every task. This skill designs the playbook that does fit — once, rigorously — and leaves it behind as a validated, rerunnable artifact.

When to use

  • a task shape recurs and no bundled skill covers it (release cut from an unusual monorepo, a compliance sweep, a data backfill dance)
  • a one-off that is high-stakes enough to deserve a designed procedure rather than improvisation

When not to use

  • a bundled skill already fits — check map first
  • work one agent finishes inside one session — just do it via execute

Design workflow

  1. Objective: one sentence, executable by a stranger with no chat access. Name the deliverable and the done-check.
  2. Steps: decompose until each step ends in a check a command can verify. A step whose exit cannot be checked is two steps, or a guess.
  3. Verification: for each step, the exact command and the expected observable. For the whole playbook, the final evidence that says done.
  4. Failure handling: for each step, what to do when its check fails — retry, rollback, or stop-and-report. Improvisation under failure is where playbooks die; write the branch.
  5. Dry-run the playbook once end to end on the real task, fixing the playbook where reality disagreed, not the other way around.

Storage and validation

Resolve the checker and the playbooks directory it validates first — from a target project the script does not exist at ./scripts/, ordinary launches may leave both plugin-root variables unset, and OMC_STATE_DIR or a parent .omc-workspace can redirect OMC state away from the checkout's literal .omc/:

checker="${CLAUDE_PLUGIN_ROOT:-${OMC_PLUGIN_ROOT:-}}/scripts/check-omc-playbooks.mjs"
if [ ! -f "$checker" ] && omc_bin="$(command -v omc 2>/dev/null)"; then
  # Portable realpath: BSD readlink has no -f, so resolve through cd/pwd -P.
  bin_dir="$(dirname "$omc_bin")"
  bin_dir="$(cd "$bin_dir" && pwd -P)"
  for d in "$bin_dir" "$bin_dir/.." "$bin_dir/../.." "$bin_dir/../../.."; do
    if [ -f "$d/scripts/check-omc-playbooks.mjs" ]; then checker="$d/scripts/check-omc-playbooks.mjs"; break; fi
  done
fi
playbooks_dir="$(node "$checker" "$PWD" --print-playbooks-dir)"

Standalone installs reach the script by walking up from the omc binary to the package root that ships scripts/. If the checker still cannot be found, report the resolution failure instead of skipping validation.

Store the result at $playbooks_dir/<name>.md with frontmatter:

---
extends: <bundled-skill-name>
when: <one sentence naming the requests this playbook serves>
---

Anchor every modification of a bundled step with **After** "<quoted step text>" / **Before** / **Replace** / **In**, quoting the bundled skill verbatim. Then run node "$checker" . to validate.

It fails when an OMC upgrade rewrites a step you anchored to. Report what it prints; never guess where an unanchored change belongs.

Rules

  • A step without a check does not ship.
  • The playbook states its own scope ceiling: what tasks it must NOT be used for.
  • Dry-run evidence is recorded in the verification ledger at OMC's state root (state/verification-ledger.tsv under the resolved state root) before the playbook is declared done.

コメント