ワンクリックで
pattern-creator
Use when: creating CSA pattern (PATTERN.md + skill + workflow.toml)
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Use when: creating CSA pattern (PATTERN.md + skill + workflow.toml)
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when: full dev cycle branch->plan->implement->review->PR->merge
Use when: executing tasks via csa run/review/debate, session mgmt
Use when: iterative review-fix loop until csa review --diff is clean
Use when: running CSA-driven code review, independent model selection
Use when: three-layer manager-employee orchestration for delegation
Use when: legacy alias, redirects to dev2merge pipeline
| name | pattern-creator |
| description | Use when: creating CSA pattern (PATTERN.md + skill + workflow.toml) |
Create CSA patterns: composable workflow definitions that orchestrate multi-step agent work through structured Markdown documents.
| Aspect | Skill (SKILL.md) | Pattern (PATTERN.md + companion skill) |
|---|---|---|
| Purpose | Static knowledge / persona | Executable workflow with steps |
| Triggered by | Claude Code auto-match | /name command or csa run --skill name |
| Has steps | No (free-form instructions) | Yes (## Step N: structure) |
| Control flow | None | ## IF, ## FOR, ## INCLUDE |
| Recursion safety | N/A | Role Detection mandatory |
| Config cascade | N/A | .skill.toml (package → user → project) |
| Composability | N/A | ## INCLUDE other-pattern |
Rule: If the workflow has more than 2 sequential steps with tool invocations, it should be a pattern, not a plain skill.
Read references/architecture.md for the full directory layout, companion skill
mechanism, symlink routing, and installation scopes.
Key concepts (always in context):
Companion Skill — patterns/<name>/skills/<name>/SKILL.md is the entry
point that Claude Code / Codex discovers. It does NOT execute the workflow
itself — it dispatches to the PATTERN.md. This prevents infinite recursion.
Role Detection — Every companion skill MUST include a "Role Detection"
section that checks whether the agent is the orchestrator (human user invoked
/name) or the executor (CSA spawned with "Use the <name> skill").
Symlink Routing — Weave creates symlinks in .claude/skills/ pointing
to the companion skill directory. The symlink can be renamed (to avoid
collisions) as long as it points to the correct companion skill. Changing
the symlink target from project-level to global-level determines which
PATTERN.md gets executed.
Before writing anything, answer:
csa (delegation) vs bash (local execution)?## INCLUDE)?${VAR_NAME})?Create the directory structure:
patterns/<name>/
├── PATTERN.md # Workflow definition (TOML frontmatter + steps)
├── .skill.toml # Agent config (tier, tools, token_budget)
├── workflow.toml # Machine-readable step+variable manifest
└── skills/
└── <name>/
└── SKILL.md # Companion skill (YAML frontmatter, entry point)
Read references/architecture.md for detailed specs of each file.
Read references/pattern-syntax.md for the full syntax reference:
frontmatter format, step structure, control flow directives, tool annotations,
tier annotations, and composition via ## INCLUDE.
Read references/companion-skill.md for the template and rules:
Role Detection block, Execution Protocol, Done Criteria, and the absolute
prohibition on csa commands inside executor mode.
[skill]
name = "<pattern-name>"
version = "0.1.0"
[agent]
tier = "tier-2-standard" # tier-1-quick | tier-2-standard | tier-3-complex
max_turns = 30 # advisory, not enforced
tools = [{ tool = "auto" }] # or explicit [{ tool = "codex" }, { tool = "claude-code" }]
[plan]
name = "<pattern-name>"
description = "One-line description"
[[plan.variables]]
name = "VAR_NAME"
# Repeat for each variable used in PATTERN.md
[[plan.steps]]
id = 1
title = "Step Title"
prompt = "What this step does"
tool = "bash" # optional: bash | csa | omit for orchestrator logic
on_fail = "abort" # "abort" | "skip" | { retry = N } | { delegate = "target" }
tier = "tier-1-quick" # optional tier override
condition = "${VAR}" # optional: only run if truthy
Project-level (default):
# Weave auto-links after install (requires source):
weave install <source-repo-or-path>
# Or manual symlink:
ln -s ../../patterns/<name>/skills/<name>/ .claude/skills/<name>
User-level (global):
ln -s /path/to/patterns/<name>/skills/<name>/ ~/.claude/skills/<name>
Renamed symlink (avoid collisions):
# Symlink name can differ from pattern name
ln -s ../../patterns/<name>/skills/<name>/ .claude/skills/my-custom-alias
The symlink name determines the trigger (/my-custom-alias), but CSA pattern
resolver follows the companion skill's actual directory name back to the pattern.
/name — should dispatch to CSAcsa run --skill name "prompt" — should read PATTERN.mdcsa commands## INCLUDE resolves correctly (if used)| File | Format | Required | Purpose |
|---|---|---|---|
PATTERN.md | TOML frontmatter + Markdown | Yes | Workflow definition |
skills/<name>/SKILL.md | YAML frontmatter + Markdown | Yes | Entry point for Claude Code |
.skill.toml | TOML | Recommended | Agent config (tier, tools, budget) |
workflow.toml | TOML | Recommended | Machine-readable manifest |