Install and use Sam's Skill Creator
Sam’s Skill Creator turns a repeatable workflow into a portable agent skill. It also improves existing skills and reviews weak triggers or incomplete execution. You receive the artifact, a dependency receipt, and the checks that actually ran.
Use it when you want to create or improve a skill. For a standalone prompt review, use Zach Prompting. For a business task, use the relevant task skill directly.
Before you start
Section titled “Before you start”You need an agent that supports skills, Node.js 18 or later, npm, Git, and network access to the declared source repositories. Codex structural validation also requires Python 3 and PyYAML. The recorded release checks used Python 3.12 and PyYAML 6.0.3.
Prepare a workflow description, an intended runtime, an output directory, and a concrete result to check. Synthetic examples are enough for a first attempt.
1. Install the skill
Section titled “1. Install the skill”Install it for Codex in the current project:
npx skills add ThatGuySam/skills --skill sams-skill-creator --agent codex --copy -yThe project installation contains .agents/skills/sams-skill-creator/SKILL.md. Add -g for a global installation. See collection installation for other clients and the bundle.
2. Restore the locked methods
Section titled “2. Restore the locked methods”From the project root, install the skill’s dependencies:
cd .agents/skills/sams-skill-creatornpm ci --ignore-scriptsnpm run skills:installnpm run preflight --silentThe last command prints a JSON receipt with status set to installed-content-verified and six dependency entries. Save it with your task evidence. If a command fails, follow troubleshooting.
These commands belong in the installed skill directory, which has its own package.json. Return to your project before starting the agent task.
3. Create a review-checklist skill
Section titled “3. Create a review-checklist skill”In Codex, send this prompt:
Use $sams-skill-creator to create a project-local Codex skill namedrelease-notes-review in .agents/skills/release-notes-review/.
It reviews supplied release notes for the affected user, changed behavior,migration steps when needed, and evidence for claims. It returns actionablefindings with quoted passages. It does not publish or rewrite the notes.
Include one synthetic fixture with an unsupported speed claim and one with abreaking change that lacks migration steps. Try the skill on both fixtures.Report the findings, dependency receipt, and any untested behavior.For a standalone Claude Code installation, use /sams-skill-creator. The collection bundle uses /sam:sams-skill-creator.
This is an example request, not a recorded evaluation result. The agent should produce release-notes-review/SKILL.md, fixtures, trial outputs, and a validation report.
4. Check the delivered result
Section titled “4. Check the delivered result”Confirm that the artifact exists at your requested path. One fixture’s findings should identify the unsupported speed claim. The other should identify missing migration steps. The trial must leave the supplied notes unchanged.
The report must separate structural validation from actual behavior. A valid SKILL.md alone does not prove that the review works. Any required check that could not run must remain an explicit gap.
Continue with task examples, the behavior and dependency reference, or verification and troubleshooting.