Skill Builder: Skill Creator Plugin and Skills for Claude Code, Codex and AI Agents

imageCI imagenpm imageTests imageSkills imageTested in imageLicense imageYouTube imageX imageLinkedIn

Skill creator plugin and skills for Claude Code, Codex and AI agents. Plan, write, check, eval, benchmark, package, and publish agent skills and plugins for every AI tool.

What is it · See it · Why · Install · How the skills fit · All skills · FAQ

npx skills add thenavidm/skill-builder

Built by Navid Moazzez.

What is Skill Builder?

Skill Builder is a skill creator plugin for Claude Code, Codex and other AI agents. It builds agent skills and whole plugins, and proves they work in every AI tool before it calls them done.

You say "build me a skill that writes our release notes" or "make a plugin like these three repos, but better", and it does the work. It plans the skills and agents first, writes each one to the Agent Skills standard, checks it in code, tests it with and without the skill on the same prompts, and measures it in Claude Code, Codex and Antigravity. Then it packages one source into every tool's plugin and publishes the repository.

It also works from what you already did. Say "turn what we just did into a skill" and it builds one from this chat, a screen recording or a brief. Correct it once, and the correction reaches every skill it touches, with the evals run again to prove nothing broke. Point it at someone else's skill, and it builds a better one from scratch and shows you the numbers, without copying a line.

Code does the validating, the counting and the statistics, so the same skill gets the same result every time. The model only judges what needs judgment, such as whether an output answered the task, and says when it did.

Ask your AI tool in plain words:

• "Build me a skill that writes our release notes from merged pull requests."
• "Turn what we just did into a skill."
• "Here's a skill someone else made. Build a better one and prove it."
• "Why doesn't my skill fire when I ask for meeting notes?"
• "Build me a plugin for our podcast: guest booking, show notes, clips."

See it

A small skill that sets a team's style for dates, tested on 19 October 2026 on two tasks, with and without the skill, in three tools:

$ skill-builder bench house-dates --runs 1
tool           with  without  change
Claude Code    100%      58%     +42
Codex          100%      25%     +75
Antigravity    100%      n/a     n/a
Antigravity: 2 runs stopped on a permission headless mode can't ask for, and are left out.

Skill Builder's own 18 skills, routed on 29 real prompts the same day: every prompt reached the right skill in Antigravity, 90% in Claude Code (the rest went to other installed skill builders, none to the wrong one of ours) and 86% in Codex.

Why Skill Builder

• Proven in the tools you use, not on one prompt. It runs your skill's prompts in Claude Code, Codex and Antigravity and tells you which ones fired it, and which other installed skill took a prompt instead (trigger).
• With and without, as a number. The same tasks run with the skill and without it, several times each, and the difference is the result. An assertion that passes as often without the skill is named, because it proves nothing.
• Checked in code against every tool's rules. Fields no tool reads (a misspelled user-invokable silently does nothing), names that don't match their folder, paths only one tool has, macOS-only commands, keys pasted into a skill. Every rule names its source and the date it was checked.
• It knows what your skills cost. Every installed description sits in context on every message, and tools drop descriptions when the list overflows. It measures each skill's share before your skills go quiet.
• Whole plugins, not only skills. The skills, the agents, the entry skill that routes between them, the repository, every tool's manifest from one source, and releases that reach every user.
• It learns. Correct it once and the correction reaches every skill it applies to, with their tests run again.
• Nothing copied. When it builds a better version of someone else's skill, it proves no run of eight words came across.
• Your AI tool, your model. The same skills install in Claude Code, Codex, Gemini CLI, Antigravity and Cursor. No agent is pinned to a model.

Compared to writing skills by hand, a skill-writing prompt, and Claude on its own

Who it's for

• Anyone who repeats a task with AI. Turn the way you do it into a skill that does it your way every time.
• Plugin makers. Build, test, package and release a plugin for every AI tool from one source.
• Teams. Skills checked and measured before they reach everyone's tools, and corrections that reach every skill.
• Anyone with too many skills installed. Find out which ones cost the most and why some stopped firing.

Install

In Claude Code:

/plugin marketplace add thenavidm/skill-builder
/plugin install skill-builder@skill-builder

In Codex: codex plugin marketplace add thenavidm/skill-builder, then codex plugin add skill-builder@skill-builder. Any other AI tool: npx skills add thenavidm/skill-builder.

Then ask: "make me a skill that writes our release notes". Every other AI tool, and how to update and remove it, is in section 2.

How the skills fit

                              skill  (the entry point: setup, health check, routing)
                                │
         ┌──────────────────────┼──────────────────────┐
         ▼                      ▼                      ▼
   skill-plan             skill-create           skill-plugin, skill-repo
   (what to build)        (one skill)            (a whole plugin and its repo)
         │                      │                      │
         └──────────────────────┼──────────────────────┘
                                ▼
     prove it   skill-check · skill-tokens · skill-description · skill-eval · skill-bench
                                │
                                ▼
     keep it    skill-improve · skill-tidy · skill-learn · skill-compete · skill-facts
                                │
                                ▼
     ship it    skill-package · skill-release · skill-publish

How you use it

Install the plugin, then ask in plain words. Your AI tool picks the right skill and runs it from start to finish.

You can also type a command. One entry point routes everything: /skill <command>.

/skill setup                     one interview, saved for every skill
/skill plan "<goal or repos>"    which skills and agents, and why
/skill create "<job>"            a skill to the standard, checked
/skill eval <path>               with and without the skill, on the same prompts
/skill bench <path>              the evals in every AI tool
/skill package                   one source into every tool's plugin

Commands are written /skill ... here. In Claude Code and Codex the full name carries the plugin's prefix, /skill-builder:skill create ....

Features

Every skill is in section 5, and all 8 agents in section 6.

Contents

1. What you can ask it

• Build me a skill that writes our release notes from merged pull requests.
• Turn what we just did into a skill.
• Here is a screen recording of how I do our weekly report. Make it a skill.
• From now on, links never go in the body of a post. Apply that to every skill it affects.
• Here is a skill someone else made. Build a better one and prove it.
• Here are three repos that do something like what I want. Tear them down and plan a better plugin: which skills, which agents, and why.
• Why does my skill not trigger when I ask for it? Fix the description and prove it.
• Run my skill's evals with and without it, three times each, and show me where it helps.
• Does this skill work in Codex and Gemini CLI too, or only in Claude Code?
• How many tokens does my plugin cost on every message?
• Turn this Claude Code skill into a plugin that installs in every AI tool, from one source.
• Publish it as a repository with a README, topics, a changelog and a release.

The fifth one is the point. A skill that only works where it was written is half a skill.

2. Install

You need one of the AI tools below. Its checks run on Node 22 or newer.

Claude Code

claude plugin marketplace add thenavidm/skill-builder
claude plugin install skill-builder@skill-builder

Or inside a session: /plugin marketplace add thenavidm/skill-builder, then /plugin install skill-builder@skill-builder. Claude Code leaves auto-update off for marketplaces it doesn't run itself, so turn it on once in /plugin (Marketplaces, skill-builder, Enable auto-update). New releases then arrive in the background when Claude Code starts.

Codex

codex plugin marketplace add thenavidm/skill-builder
codex plugin add skill-builder@skill-builder

Or run codex, type /plugins, and install Skill Builder from the marketplace you added. Start a new session before you use it.

Gemini CLI

gemini extensions install https://github.com/thenavidm/skill-builder --auto-update

Antigravity

agy plugin install https://github.com/thenavidm/skill-builder

That is the Antigravity CLI. For the Antigravity app, copy the repository folder into ~/.gemini/config/plugins/ for every workspace, or into .agents/plugins/ for one project. Antigravity gets the skills.

Cursor

Copy the repository folder to ~/.cursor/plugins/local/skill-builder and run Developer: Reload Window. Copy it rather than linking it: Cursor skips a link there that points outside that folder.

Any tool the skills CLI supports (Windsurf, OpenCode, GitHub Copilot, Cline and others), the skills only:

npx skills add thenavidm/skill-builder

Claude Desktop and claude.ai (Pro, Max, Team and Enterprise plans)

1. Open Customize in the left sidebar, then the Plugins tab.
2. Click Add, choose Add marketplace, then Add from a repository.
3. Enter thenavidm/skill-builder and install Skill Builder from it.

The whole plugin runs in Cowork: the skills, the agents. In chat, on the web or in the desktop app's Chat tab, the skills run, but agents do not.

ChatGPT and the Codex app

Plugins install from the Plugins tab. To offer Skill Builder there, a workspace admin imports thenavidm/skill-builder as a GitHub marketplace for the team. On your own account, add it with the Codex commands above.

A terminal

npm install -g @thenavidm/skill-builder
skill-builder check path/to/your/skill

That is the engine on its own, for CI or scripts: check, tokens, trigger, eval, bench, package, release and the rest, with skill-builder --help for the list. You don't need it for the skills: they run it through npx -y @thenavidm/skill-builder@latest, so they always use the newest release.

What each tool runs

Where a tool has no agents, the skills run the same steps one after another in your main conversation and reach the same result. Only the time differs.

Updating and removing

Every release is listed in CHANGELOG.md, and each one goes out to all of these at once: the marketplaces read this repository, and npm gets the same version.

3. Set up

Run /skill setup once in your project. It asks a short interview and writes the answers to .agents/skill/setup.md, which every skill reads before it starts.

.agents/skill/setup.md is plain Markdown in your project. Edit it by hand any time.

4. How a skill is built

1. Plan. From a goal, from what you just did (this chat, a screen recording, a pasted brief), or from example repos torn down side by side: what each does, what is missing, which skills and agents a better one needs, and which jobs code should do instead of the model.
2. Write. Each skill to the Agent Skills standard: a name that says what it does, a description that leads with the words people use, steps, and references loaded only when a step needs them.
3. Check, in code. Only the fields every tool reads, the name matching its folder, YAML that reads the same in every parser, no paths that exist in only one tool, every reference file present, and no number without a source.
4. Test. The same prompts with and without the skill, several runs each, graded against written assertions, with a page that shows you every output.
5. Measure in every tool. The same evals in Claude Code, Codex, Gemini CLI and Cursor, with tokens, time and how often the results agree.
6. Package. One product.json builds every tool's manifest, so nothing is copied by hand and nothing drifts.
7. Learn. Every correction you give is applied to each skill it touches, after your yes on each change, and their evals run again. A lesson that reaches beyond skills goes to your AI tool's memory instead.
8. Prove. Against a plain AI and against the skill you wanted to beat, on the same tasks and the same model, with the numbers in the README.

5. Skills

Skill Builder has 24 skills, 18 shipped and 6 in the next release, each a folder in skills/ with a SKILL.md your AI tool loads only when the job needs it. Every skill also works without agents, so results are the same in every AI tool. Skills marked Release 1 are in 0.1.0; Release 2 comes next.

Start and plan

Write

Check and test

Improve and ship

6. Agents

An agent is a helper your AI tool starts for one job. It gets its own instructions and a fresh context, does the job, writes its result to a file, and reports back in a few lines. Your main conversation stays short, and several agents can work at the same time. No agent is pinned to a model: each runs on the one you already use.

1. A skill starts it, at a fixed step. This is the normal way. You never have to name an agent.
2. Your AI tool starts it on its own, when what you ask matches the agent's description.
3. You ask for it by name. In Claude Code you can also type @ and pick it.

Codex has no agents in plugins, so there the skill does each agent's steps itself and writes the same file. A test fails if an agent exists that no skill starts.

What an eval runs

/skill eval skills/release-notes
  reads the skill, its assertions and its prompts
  at the same time, for every prompt and every run:
    skill-executor (with the skill)       -> runs/<prompt>-with-<n>/
    skill-executor (without the skill)    -> runs/<prompt>-without-<n>/
  skill-grader                            -> grades-<date>.json
  skill-analyzer                          -> benchmark-<date>.json
  the review page

The agents

7. Evidence

Every platform rule, limit and number a skill relies on lives in facts/, with its source, the date that source was last checked, and how strong the evidence is. A skill points at the fact; it never states an undated claim. Each release checks every source again.

A result is never stronger than the weakest fact behind it.

The format rules it checks

Checked against each tool's own documentation on 2026-10-13.

Ask for the evidence behind anything: /skill facts description length or /skill facts codex agents.

How Skill Builder proves itself

Every release runs the same tasks through Skill Builder, a plain AI with no plugin, and other skill-creation tools, on the same model:

1. A skill built from a described workflow.
2. One correction spread across three skills.
3. A published skill beaten.

The results, with the date and the versions, go in this section at each release.

8. Your data

Nothing is uploaded anywhere. Evals run in your own AI tool, on your own account.

There is no telemetry and no backend of ours.

9. Notes and gotchas

• Evals cost tokens. Every prompt runs with and without the skill, several times. The skill shows the estimate first.
• Codex has no agents in plugins. A skill that needs an agent there does the agent's steps itself, so skill-bench measures both paths.
• A description that names a phrase can make a model use it. Descriptions are written as what the skill does and when, not as lists of forbidden words.
• Many skills crowd the listing. With a lot installed, a tool may shorten the descriptions it lists, so every description leads with the words people use, and skill-tokens shows the total.
• A green check is not a working skill. skill-check proves the shape; only skill-eval proves the skill does its job.

10. Troubleshooting

Run /skill doctor first. It checks Node, the scripts and which AI tools it can run, and names the fix for anything missing.

Limitations

• Gemini CLI refuses free personal Google accounts (October 2026), so its tests need a paid or API-key account. Antigravity, Google's replacement, is tested.
• Cursor installs the plugin, but its CLI isn't part of the automated tests yet.
• Windows runs every check; running prompts in each tool is tested on macOS so far.
• A crowded skill list. With hundreds of skills installed, Claude Code and Codex list most of them by name only. Skill Builder measures it, but can't change the tool's budget for you.
• Six skills are next, not shipped: agents, from-docs, review, library, convert and update.

11. FAQ ❓

What is Skill Builder?

Skill Builder is a plugin that builds agent skills and whole plugins, and proves they work. It plans them, writes them to the Agent Skills standard and checks them in code against each AI tool's own rules. It tests them with and without the skill in Claude Code, Codex and Antigravity. Then it sets up, packages and releases the repository. It installs in Claude Code, Codex, Gemini CLI, Antigravity and Cursor.

What is a plugin, a skill and an agent?

A plugin is a bundle you install into your AI tool. A skill is a set of instructions and tools the AI tool loads only when a job needs it. An agent is a helper the AI tool can run in parallel for part of a job, such as grading outputs while others run the prompts.

Why doesn't my skill trigger?

There are three causes, in the order we see them.

1. The list of skills is full. Claude Code fits every skill's name and description into 1% of the model's context window, and Codex into 2%. Past that, descriptions are dropped (in Claude Code, the least-used skills' first), and a skill without its description can't be matched to a request. On a 200,000-token model, one 27-skill plugin we measured needs 168% of Claude Code's budget on its own. /skill tokens shows what yours cost.
2. The name says nothing. In our tests, clear names routed every prompt to the right skill in Claude Code, Codex and Antigravity, even with vague descriptions. With vague descriptions and names like helper-one, routing fell to 88%, 63% and 50%.
3. The description leads with the wrong words. /skill description tests it on prompts that should fire it and prompts that shouldn't, in each tool, and rewrites it until it does.

How many skills can I install before they stop working?

There is no fixed number. It depends on the model's context window and on how long every installed description is. A model with a 1,000,000-token window has five times the room of a 200,000-token one. /skill tokens shows each skill's cost on every message and the whole set's share of each tool's budget, so you can see the list fill before skills go quiet.

I set a field in my skill and nothing changed. Why?

Claude Code skips a frontmatter field it doesn't know, without an error. A field no tool reads, such as auto-activate, or a misspelling, such as user-invokable, does nothing and nothing tells you. On one computer we found that misspelling in 32 skills. /skill check names every field no tool reads, and what to use instead.

How do I make a skill run only when I ask for it?

Each tool has its own switch: disable-model-invocation: true in SKILL.md for Claude Code and Cursor, and allow_implicit_invocation: false in agents/openai.yaml for Codex. Write the first; /skill package writes the second for you. Gemini CLI and Antigravity have no switch yet, so for them the description says when the skill is for.

Will a skill I wrote for Claude Code work in Codex, Gemini CLI or Antigravity?

Usually it does, as long as it avoids what only one tool has: ${CLAUDE_SKILL_DIR}, $ARGUMENTS, commands that run before the skill loads, folders such as ~/.claude/, and commands that exist on macOS only. /skill check finds each one and says what works everywhere. /skill bench then runs the skill's evals in every tool you have, so "it works in Codex too" is a number.

Is it safe to run the tests?

Each run gets its own throwaway folder and can write only there: Codex in its workspace-write sandbox, Claude Code with edits accepted, Antigravity in its accept-edits mode. Skill Builder never switches permission checks off. A run that stops because headless mode can't ask for a permission is marked blocked and left out of the numbers, so it can't pass for a failure of the skill.

The runs can see the rest of your computer. A run without the skill may find your other skills, which keeps the comparison true to what you would see.

How are the outputs graded, and can a grader get it wrong?

Code grades everything code can (a file is written, contains something, matches a pattern, is valid JSON), so those grades come out the same on every run. A model grades only what needs judgment, and its grade counts only if it quotes the output and the quote is really there. Assertions that pass as often without the skill as with it are named, because they prove nothing about the skill.

What does it cost?

Skill Builder is free and MIT licensed. The tests run in your own AI tool, on your own account, so they use its tokens. Every command that runs prompts says how many runs it will make before it starts, and a trigger test stops each run as soon as it knows which skill fired: 24 runs across three tools take about a minute.

Can it build a whole plugin and set up the repository?

Yes, it can do both. /skill plugin plans and writes the skills, the agents and the entry skill that routes to them. /skill repo writes everything a published plugin needs: the README as its specification, a licence, a changelog, how to report a bug and a vulnerability, CI, search topics, and one product.json that builds every tool's manifest. A repository it writes passes Claude Code's, Gemini CLI's and Antigravity's own validators from the first commit.

How do I release an update so people actually get it?

Every tool decides that a copy is out of date from the version number, so a release that keeps the old number reaches no one. /skill release bumps the version, moves the changelog notes under it, rebuilds every manifest, then lists the tag, the push, the GitHub release and npm, each for your yes.

Someone sent me a skill. Can I check it before I install it?

Yes, you can. /skill check reads it without running anything: fields that do nothing, files it points to that aren't there, commands for one system only, and secrets. Shared skill folders sometimes carry a .env file with live keys. Check fails that, and keys pasted into files, naming the file and the kind of key, never the key itself.

Does it work on Windows?

Its checks are plain Node and run wherever Node 22 does. Running prompts in each AI tool is tested on macOS so far; Windows is next.

Does it send my skills anywhere?

No, it doesn't. Nothing is uploaded and there is no telemetry. Everything runs on your computer, in your own AI tool.

How do I remove it?

Uninstall the plugin in your AI tool (claude plugin uninstall skill-builder@skill-builder in Claude Code), and delete .agents/skill/ if you want your setup and results gone too.

Versions

See CHANGELOG.md.

Questions

Run into a problem or have a question? Open an issue and I will help.

Built by

Navid Moazzez

Navid Moazzez is a leading AI business strategist, and the host of the AI Creator Summit, watched by 100,000+ creators. He helps creators and founders master AI and build their own AI Operating System (AI OS) to automate their business and life. He creates useful free tools, MCP servers and CLIs for their own workflows.

License

MIT. Free to use, modify, and share.

Not affiliated with, endorsed by, or connected to Anthropic, OpenAI, Google, Cursor, or any other company named here. Their names identify the services this works with.

© 2026 Skill Builder™ is a trademark of Navid Media. All rights reserved. Made with ❤️ by Navid Moazzez.