Agent Patterns
Use this skill to turn repeatable agent workflows into first-class skills. A pattern is not a prompt: it must have a trigger, scope, workflow, inputs, outputs, tool permissions, human gates, failure modes, and verification.
SKILL.md is the workflow source of truth. A pattern-manifest.json is only a lightweight index for search, validation, and registry views; it must not duplicate workflow steps, failure modes, inputs, outputs, or gates.
Route
- Search first for existing skills, pattern manifests, and similar workflows before creating a new one.
- Choose the smallest useful shape:
- One skill when the workflow has a single clear trigger and procedure.
- Multiple skills when patterns have different triggers, risk profiles, or verification loops.
- A meta skill plus concrete skills when building a reusable pattern pack.
- Write
SKILL.md as the product surface. Put deterministic helpers under scripts/; put only lightweight manifests or schemas under references/.
- Keep scripts generic. They may validate, list, search, or show manifests, but they must not be the only usable interface.
- Validate each skill folder with the skill creator
quick_validate.py script.
Pattern Contract
Every concrete pattern skill should define in SKILL.md:
- Trigger and non-trigger in the frontmatter description.
- Required inputs and available evidence.
- Ordered workflow steps.
- Tool permissions and human gates.
- Failure modes with explicit stop/report behavior.
- Done-when and verification commands.
When the workflow uses high-risk operations such as shell, network, secrets, external comments, writes, payments, or destructive actions, require a human gate before that step.
Each concrete pattern skill may include references/pattern-manifest.json with only these index-level facts: id, skill, title, summary, category, tags, maturity, risk level, capabilities, whether a human gate is required, source, and maintenance mode.
Resources
references/pattern-manifest.schema.json: schema for lightweight pattern manifests.../*/references/pattern-manifest.json: per-skill manifests, each pointing back to its owning SKILL.md.scripts/pattern_tool.py: helper for listing, searching, validating, and showing pattern manifests.
Use the helper from the skill directory root:
python3 skills/agent-patterns/scripts/pattern_tool.py list
python3 skills/agent-patterns/scripts/pattern_tool.py search "pr review"
python3 skills/agent-patterns/scripts/pattern_tool.py validate
python3 skills/agent-patterns/scripts/pattern_tool.py show pr-review-risk-plan.v1
python3 skills/agent-patterns/scripts/pattern_tool.py show skill-lifeguard.v1
python3 skills/agent-patterns/scripts/pattern_tool.py show review-gate.v1
Output Rules
- Prefer a skill folder over a standalone script.
- Do not create README, quick reference, changelog, or installation docs inside a skill folder.
- Do not put full workflow JSON in the meta skill. Keep workflows in concrete
SKILL.md files.
- Do not mark a community or untested pattern as verified.
- Do not hide unsupported exporter behavior; report missing target support explicitly.
- Keep machine-facing IDs in English and stable across exports.
1---2name: agent-patterns-23description: Create, convert, audit, and package executable AI agent workflow patterns as Codex/Claude-style skills. Use when the user wants an agent workflow library, reusable workflow schema, pattern-to-skill conversion, skill pack design, exporter/validator behavior, curated coding-agent patterns, or wants to avoid prompt-library or hardcoded-script designs.4---56# Agent Patterns78Use this skill to turn repeatable agent workflows into first-class skills. A pattern is not a prompt: it must have a trigger, scope, workflow, inputs, outputs, tool permissions, human gates, failure modes, and verification.910`SKILL.md` is the workflow source of truth. A `pattern-manifest.json` is only a lightweight index for search, validation, and registry views; it must not duplicate workflow steps, failure modes, inputs, outputs, or gates.1112## Route13141. Search first for existing skills, pattern manifests, and similar workflows before creating a new one.152. Choose the smallest useful shape:16 - One skill when the workflow has a single clear trigger and procedure.17 - Multiple skills when patterns have different triggers, risk profiles, or verification loops.18 - A meta skill plus concrete skills when building a reusable pattern pack.193. Write `SKILL.md` as the product surface. Put deterministic helpers under `scripts/`; put only lightweight manifests or schemas under `references/`.204. Keep scripts generic. They may validate, list, search, or show manifests, but they must not be the only usable interface.215. Validate each skill folder with the skill creator `quick_validate.py` script.2223## Pattern Contract2425Every concrete pattern skill should define in `SKILL.md`:2627- Trigger and non-trigger in the frontmatter description.28- Required inputs and available evidence.29- Ordered workflow steps.30- Tool permissions and human gates.31- Failure modes with explicit stop/report behavior.32- Done-when and verification commands.3334When the workflow uses high-risk operations such as shell, network, secrets, external comments, writes, payments, or destructive actions, require a human gate before that step.3536Each concrete pattern skill may include `references/pattern-manifest.json` with only these index-level facts: id, skill, title, summary, category, tags, maturity, risk level, capabilities, whether a human gate is required, source, and maintenance mode.3738## Resources3940- `references/pattern-manifest.schema.json`: schema for lightweight pattern manifests.41- `../*/references/pattern-manifest.json`: per-skill manifests, each pointing back to its owning `SKILL.md`.42- `scripts/pattern_tool.py`: helper for listing, searching, validating, and showing pattern manifests.4344Use the helper from the skill directory root:4546```sh47python3 skills/agent-patterns/scripts/pattern_tool.py list48python3 skills/agent-patterns/scripts/pattern_tool.py search "pr review"49python3 skills/agent-patterns/scripts/pattern_tool.py validate50python3 skills/agent-patterns/scripts/pattern_tool.py show pr-review-risk-plan.v151python3 skills/agent-patterns/scripts/pattern_tool.py show skill-lifeguard.v152python3 skills/agent-patterns/scripts/pattern_tool.py show review-gate.v153```5455## Output Rules5657- Prefer a skill folder over a standalone script.58- Do not create README, quick reference, changelog, or installation docs inside a skill folder.59- Do not put full workflow JSON in the meta skill. Keep workflows in concrete `SKILL.md` files.60- Do not mark a community or untested pattern as verified.61- Do not hide unsupported exporter behavior; report missing target support explicitly.62- Keep machine-facing IDs in English and stable across exports.
majiayu000