ryderfreeman4logos/pattern-creator

Pattern Creator

Create CSA patterns: composable workflow definitions that orchestrate multi-step agent work through structured Markdown documents.

Patterns vs Skills — When to Use Which

| 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.

Pattern Architecture (Progressive Loading)

Read references/architecture.md for the full directory layout, companion skill mechanism, symlink routing, and installation scopes.

Key concepts (always in context):

1. 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.

2. 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").

3. 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.

Creation Process

Step 1: Define Scope

Before writing anything, answer:

1. What problem does this pattern solve?
2. What are the steps (draft a numbered list)?
3. Which steps need csa (delegation) vs bash (local execution)?
4. Does it compose with existing patterns (## INCLUDE)?
5. What variables does it need (${VAR_NAME})?

Step 2: Scaffold

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.

Step 3: Write PATTERN.md

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.

Step 4: Write Companion Skill

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.

Step 5: Write .skill.toml

[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" }]

Step 6: Write workflow.toml

[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

Step 7: Install

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.

Step 8: Test

1. Invoke as orchestrator: /name — should dispatch to CSA
2. Invoke as executor: csa run --skill name "prompt" — should read PATTERN.md
3. Verify Role Detection: executor must NOT run csa commands
4. Verify ## INCLUDE resolves correctly (if used)

Quick Reference

| 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 |