Skill workflows

Select a skill because the task needs its procedure

A SKILL.md procedure may include scripts, references, templates, and assets. It does not install software, authenticate a service, or widen task authority.

Need	Route	Evidence
unfamiliar repository/runtime	env-reconciler	checked environment plan
browser UI	browser-workbench or Playwright	screenshots, trace, console/network evidence
native macOS app	desktop-workbench	live observation and reopened result
sustained/remote work	long-running-work + df-task	checkpoint, logs, artifacts
game/asset	router, then engine workflow	play slice and imported/rendered evidence
current APIs/libraries	research or researcher	dated primary sources and marked inference

Read the selected skill fully. Keep project conventions in project instructions and reusable procedure detail in skills.

Discover before duplicating

Use find-skills when asking whether a capability exists. If the name/source is known, inspect it with:

rg --files ~/.claude/skills ~/.cursor/skills ~/dotfiles/home/dot_claude/skills

Do not create a broadly named skill for a one-off selector or project-local source list. A reusable skill needs a stable input/output contract.

Review a change with Ponytail

ponytail-review checks a change and its connected callers for bugs, security, load, missing tests, performance, and unnecessary code. The installer declares only this skill in agent-skills.txt and shares it through the existing skill tree. It requires the managed Node/npm setup; install missing declared skills with bash install/skills-sync.sh sync.

In a fresh Codex chat, send:

$ponytail-review Review my uncommitted changes.

Name staged changes, a branch, a PR, or files to change the scope. With no scope, the skill reviews uncommitted changes, or the last commit when there are none. Expect ranked findings with file locations, concrete failure cases, and proposed fixes. The review changes no code; its result only establishes the checks it actually performed. The agent may also select it for review requests. It does not install Ponytail’s coding mode, other skills, or lifecycle hooks.

Upstream: review procedure and skills-only installation. The authoritative installer and reviewed content baseline are skills-sync.sh and agent-skills.lock.json.

Author at the correct ownership layer

Use skill-creator for a new Codex skill. Decide ownership first:

Repository-controlled reusable procedure -> home/dot_claude/skills/<name>/SKILL.md
Upstream/community installed procedure    -> packages/agent-skills.txt

Never assign both. Keep unique project methods in a project AGENTS.md, script, or README.

Write a procedure that can be checked

Include trigger, initial observation, source of truth, steps, prerequisite diagnostics, and evidence boundary:

## Before editing
Read the schema and identify the owned fixture directory.
## Procedure
Run the repository normalizer, then validate every changed fixture.
## Evidence boundary
Parsing confirms schema conformance; it does not prove remote acceptance.

Do not embed credentials, uncontrolled recursive deletion, or unverified product claims.

Validate behavior, not prose alone

Run representative isolated tasks with fixed inputs, model, effort, and tools; retain artifacts plus an independent check. Use writing-skills when a baseline comparison is warranted. Parsing Markdown proves shape only. For Codex capability assumptions, consult official skills documentation and test the local client after installation.