Skip to content

feat: Claude Code Routines パイプライン prompt 6種を追加#124

Open
TakuTsuchida wants to merge 7 commits into
shanraisshan:mainfrom
TakuTsuchida:feat/routines-pipeline-prompts
Open

feat: Claude Code Routines パイプライン prompt 6種を追加#124
TakuTsuchida wants to merge 7 commits into
shanraisshan:mainfrom
TakuTsuchida:feat/routines-pipeline-prompts

Conversation

@TakuTsuchida

Copy link
Copy Markdown

背景 / 課題

TakuTsuchida/claude-code-scheduler#10 のコメントで設計された、technical-fundamental-backtester 向けラベル駆動の自律開発パイプライン(requirement → user-story → design → implement → PR → review/merge)を、コスト面(ECS + Anthropic API はサブスク枠と別課金)からクラウドの Claude Code Routines(サブスク枠で実行)として運用したい。

これまで上流(要件定義・設計)と品質(バグ調査)の routine prompt が存在せず、パイプラインが下流(実装・レビュー)の launchd 方式しか揃っていなかった。

目的・スコープ

目的: クラウド Routines に登録する 6 種の prompt を、版管理可能な登録用ソースとして整備する。
対象外:

  • クラウド Routines の実登録・実行確認(UI 操作のため別途)
  • 既存 launchd 方式(.claude/scripts/routine-engineering.sh01.Implementation.md / 02.pr-review.md)の変更
  • スケジュール最適化(実行上限 Pro=5/day, Max=15/day を踏まえた本数調整)

変更概要

.claude/scripts/prompts/routines/ を新設し、issue 記載の全タスクを 1 ファイル 1 routine で作成(各ファイル先頭に Model / Schedule / Trigger / GitHub Repository を記載):

  • 01.create-user-story.md(PO / Opus / requirementuser-story
  • 02.design-from-user-story.md(architect / Opus / user-storydesign
  • 03.decompose-design-to-issues.md(architect / Opus / designimplement sub-issue 複数・各200行以内)
  • 04.implement-and-create-pr.md(engineer / Sonnet / implement→PR)
  • 05.pr-review-and-merge.md(tech-lead / Sonnet / open PR→レビュー→squash merge)
  • 06.investigate-bugs-create-issues.md(tech-lead / Opus / コードベース→implement fix issue 複数)
  • README.md(登録手順 + ルーチン一覧 + 実行上限の留意点)

期待される効果

  • パイプライン全工程がクラウド Routines として登録可能になり、サブスク枠で自律運用できる
  • prompt が版管理され、モデル/スケジュール/ラベル体系が一元的に可視化される
  • 04/05 を issue 指定どおり Sonnet に固定し、上流/品質を Opus にすることでコストを最適化

品質チェック

  • 6 routine が issue コメントの全タスクと 1:1 対応
  • 各ファイルに Model(Opus/Sonnet)/ Schedule / Trigger / Repository を記載
  • ラベル体系を requirement/user-story/design/implement に統一、各 prompt は自己完結
  • クラウド Routines への実登録・実行確認(UI 操作のため別途)

レビュー観点

  • ラベル体系を既存 launchd の routines: 名前空間ではなく issue 準拠の素ラベルに統一した点
  • 各 routine の Process 手順・成果物・スコープ制約が役割(PO/architect/engineer/tech-lead)と整合しているか

参照: TakuTsuchida/claude-code-scheduler#10

Taku Tsuchida and others added 7 commits June 21, 2026 23:05
Cloud Routine source for claude.ai/code/routines: the PO persona turns the
oldest `requirement` issue into a `user-story` issue so the autonomous
pipeline has a value-framed, acceptance-criteria-backed starting point.

Co-Authored-By: Claude <noreply@anthropic.com>
Cloud Routine source: an architect persona converts the oldest `user-story`
issue into a `design` issue. Centralizing extensibility decisions here keeps
downstream implementation issues small and consistent with existing patterns.

Co-Authored-By: Claude <noreply@anthropic.com>
Cloud Routine source: splitting the `design` issue into <=200-line `implement`
sub-issues is what keeps each implementation run cheap enough for Sonnet and
each PR small enough to review autonomously.

Co-Authored-By: Claude <noreply@anthropic.com>
Cloud Routine source adapted from the launchd 01.Implementation prompt, but
self-contained and pinned to Sonnet per the pipeline cost model, and unified
on the `implement` label so it chains cleanly after design decomposition.

Co-Authored-By: Claude <noreply@anthropic.com>
Cloud Routine source adapted from the launchd 02.pr-review prompt, pinned to
Sonnet and made self-contained. Gating merge on green CI is what lets the
review stage run unattended without lowering the quality bar.

Co-Authored-By: Claude <noreply@anthropic.com>
Cloud Routine source: a weekly tech-lead sweep scoped by perspective
(frontend / backend / contract / screens / reuse) rather than crawling all
code, emitting `implement`-labeled fixes so quality work re-enters the same
implement->PR->merge loop.

Co-Authored-By: Claude <noreply@anthropic.com>
…ing it

Routines are UI-registered, not file-loaded, so the source files alone leave
the Model/Schedule/Repository/Trigger settings implicit. This index makes the
registration steps and the per-run quota caps explicit for whoever sets them up.

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant