agentleFS
garbageekgarbageek/claude-plugins/skills/create-skill/SKILL.md

create-skill

Skill1 starsChanged 40 days agoMITView on GitHub

Create, edit, evaluate, or optimize a reusable Agent Skill (SKILL.md) for Claude Code, Codex, or both. Use for standalone skills, skill instructions inside an existing plugin, skill trigger descriptions, evaluations, benchmarks, and .skill packaging. Route whole-plugin creation and release management to create-plugin or update-plugin instead.

149 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use review-bugbot, migrate-to-skills and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • 0mlkrizzz655335vSkill

    review-bugbot

    缺陷导向代码审查。用 Bugbot 子代理找 bug。

    Also in 5 projects that use this

  • 32103210Skill

    migrate-to-skills

    Convert 'Applied intelligently' Cursor rules (.cursor/rules/*.mdc) and slash commands (.cursor/commands/*.md) to Agent Skills format (.cursor/skills/). Use when you want to migrate rules or commands to skills, convert .mdc rules to SKILL.md format, or consolidate commands into the skills directory.

    Also in 8 projects that use this

  • 32103210Skill

    update-cursor-settings

    Modify Cursor/VSCode user settings in settings.json. Use when you want to change editor settings, preferences, configuration, themes, font size, tab size, format on save, auto save, keybindings, or any settings.json values.

    Also in 8 projects that use this

  • PandelisZSkill

    update-cli-config

    View and modify Cursor CLI configuration settings in ~/.cursor/cli-config.json. Use when the user wants to change CLI settings, configure permissions, switch approval mode, enable vim mode, toggle display options, configure sandbox, or manage any CLI preferences.

    Also in 7 projects that use this

  • yawningwhaleAISkill

    build-premium-website

    Build a premium, animated marketing website (React + Vite + Tailwind + GSAP) for any industry. Use when the user asks to "build a website", "make a landing page for [business]", "build a marketing site", or wants a high-end animated company website. Adapts color, copy, services, and signature animation to the industry.

    Also in 8 projects that use this

  • yawningwhaleAISkill

    instantly-campaign

    Create a new Instantly cold-email campaign in the proven 5-email structure of the user's top-performing campaign (~16% reply rate). Use when the user wants to launch a new outbound campaign, write a cold email sequence, create an Instantly campaign, draft outreach for a new service or niche, or set up a sales sequence.

    Also in 7 projects that use this

  • +3 more from yawningwhaleAI

The file

Show its rules and text

We split this skill into the rules it gives an agent

14 rules. 0 are common, 6 are distinctive, and 8 are very specific to this project.

  • 0

    Common

    Most of the 139 edited versions we read keep these.

  • 6

    Distinctive

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read and apply the delivery contract within this skill's scope and authorization.

    line 8
  2. Very specific:

    Create or improve Agent Skills as independent, reusable workflows.

    line 10
  3. Distinctive:

    Improve an existing skill: read the current skill and referenced support files first.

    line 26
  4. Distinctive:

    Install or enable a skill/plugin: follow install-agent-plugins if installation is requested.

    line 33
  5. Distinctive:

    If an upstream skill, archive, or repository is supplied, adapt that source.

    line 48
  6. Very specific:

    Use a single root SKILL.md per skill, with YAML frontmatter:

    line 54
  7. Very specific:

    Write all trigger criteria in description, keeping it concrete and distinct from neighboring skills.

    line 64
  8. Very specific:

    Use references/ for substantive secondary detail, scripts/ for useful repeatable operations, assets/ for resources used in outputs.

    line 73
  9. Very specific:

    Use the supplied portable helpers from this skill's directory (SKILL_DIR):

    line 81
  10. Very specific:

    Do not package unrelated plugin files into a standalone .skill artifact.

    line 92
  11. Distinctive:

    Run representative tasks with the skill, and optionally a baseline without it or with the prior version.

    line 105
  12. Distinctive:

    Do not invent grades, token metrics, or passing evaluations.

    line 112
  13. Very specific:

    Use --host claude for the Claude CLI, or --dry-run to inspect planned runs without invoking an agent.

    line 123
  14. Very specific:

    Only aggregate after grading.json exists for runs to be scored.

    line 143

“Common” is measured: kept by at least 6 in 10 of the 139 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Create Skill

Read and apply the delivery contract within this skill's scope and authorization.

Create or improve Agent Skills as independent, reusable workflows. The user selects the task and delivery; this skill owns the SKILL.md, supporting files, and optional evidence-driven iteration. Do not turn skill-only requests into a new plugin, host installation, or hosted release.

This skill adapts the user-supplied Apache-2.0 skill-creator to the shared Claude Code and Codex skill layout. It preserves the supplied evaluation/report helpers, but does not claim the Claude CLI trigger evaluator can measure Codex skill discovery. Read host distinctions when choosing a runner.

Scope and routing

  • Create a standalone skill: produce <skill-name>/SKILL.md with resources only when needed, or place it at skills/<skill-name>/SKILL.md inside the selected existing toolkit. Default an unspecified target to both Claude and Codex, subject to the actual host/tool capability.
  • Improve an existing skill: read the current skill and referenced support files first; preserve its name, unrelated behavior, and invocation boundaries. Record a baseline snapshot only if comparative evaluation is requested.
  • Create or replace a plugin: route package design to create-plugin and source/release updates to update-plugin. Creating one skill does not authorize publishing/releasing an entire plugin.
  • Install or enable a skill/plugin: follow install-agent-plugins if installation is requested. Source packaging is not proof of installation or discovery.

1. Capture the job

From the request and supplied material, determine:

  1. What job the skill performs and what differentiates it from existing skills.
  2. Triggering language: when it should run, including near-misses and exclusions.
  3. Inputs, dependencies, actions, constraints, output files, and failure behavior.
  4. Target host(s), filesystem destination, and whether the user wants evaluation, packaging, or installation. Use existing details; ask only when the missing choice materially changes the artifact.

If an upstream skill, archive, or repository is supplied, adapt that source; do not replace it with a generic template. Do not silently introduce project names, credentials, or private workflow specifics into an otherwise reusable skill.

2. Write a focused SKILL.md

Use a single root SKILL.md per skill, with YAML frontmatter:

---
name: example-skill
description: Use when the user needs to ...; creates/updates ... .
---

The directory and frontmatter name must match; prefer lowercase kebab-case. Write all trigger criteria in description, keeping it concrete and distinct from neighboring skills. The body must explain what to do and why, with:

  • Authoritative inputs, where to find them, and actual available tools.
  • Sequenced actions and decisions, not merely a capabilities list.
  • Exact expected deliverables and preservation constraints.
  • Failure, missing-data, unsupported-host, and missing-tool behavior.
  • Links to supporting resources by valid relative paths.

Use references/ for substantive secondary detail, scripts/ for useful repeatable operations, assets/ for resources used in outputs. Avoid empty scaffolding, duplicated host-specific skill trees, and automatic side effects. Skills should remain discoverable and actionable without needing to read every bundled file. Follow the existing package's manifest authority and conventions.

3. Validate and optionally package

Use the supplied portable helpers from this skill's directory (SKILL_DIR):

cd "$SKILL_DIR" && python -m scripts.quick_validate /absolute/path/to/example-skill
cd "$SKILL_DIR" && python -m scripts.package_skill /absolute/path/to/example-skill /absolute/path/to/output

quick_validate.py and package_skill.py require Python 3 and PyYAML; they validate portable metadata and create a ZIP-format .skill archive. They are structural checks, not runtime discovery or proof of correct behavior. Within an existing plugin, also run its actual packaging validation if available. Do not package unrelated plugin files into a standalone .skill artifact.

4. Optional real evaluations and iteration

Skip evaluations if they were not requested or would incur an unwanted runtime or billing cost. For changes with measurable outcomes, reuse existing evaluation cases rather than inventing a test suite. Cases can use the supplied case schema: evals/<case-name>/prompt.md, case.yaml, optional graders/, and files/. Avoid writing assertions for purely subjective quality; use human review instead.

A measured iteration is:

  1. Run representative tasks with the skill, and optionally a baseline without it or with the prior version. Keep input prompts and files identical.
  2. Store the outputs and exact execution evidence in a separate workspace; never overwrite the source skill just to create an evaluation run.
  3. Review outputs first. If evidence supports an improvement, change the smallest relevant part of the instructions or helper. Compare again only on request.
  4. Report what was actually run, outcome differences, unknowns, and costs. Do not invent grades, token metrics, or passing evaluations.

Cross-host functional evaluation: scripts/run_task_evals.py can run cases with an installed claude or codex CLI. It creates isolated workspaces with and optionally without the skill and writes reviewable outputs, without claiming to measure automatic trigger accuracy. This command executes real model runs:

cd "$SKILL_DIR" && python -m scripts.run_task_evals --host codex --skill-path /absolute/path/to/example-skill --workspace /absolute/path/to/workspace --baseline

Use --host claude for the Claude CLI, or --dry-run to inspect planned runs without invoking an agent. This runner requires the selected authenticated CLI and a filesystem supporting its project skill layout; inspect the tool's help and the host's permissions before real calls. Compare only evidence-backed results. The runner does not itself grade outputs.

Claude-specific trigger evaluation: The preserved scripts/run_eval.py, run_loop.py, and improve_description.py call claude -p and inspect Claude skill invocation. Run them only on Claude Code when the executable and needed permissions exist and the user explicitly asks for trigger optimization. See host distinctions. These scripts must not be used to claim a Codex trigger score; Codex's skill selection needs host-native evidence.

For completed runs, the original cross-host-capable review tools are available:

python "$SKILL_DIR/eval-viewer/generate_review.py" /absolute/path/to/workspace/iteration-1 --static /absolute/path/to/review.html
cd "$SKILL_DIR" && python -m scripts.aggregate_benchmark /absolute/path/to/workspace/iteration-1 --skill-name example-skill

Only aggregate after grading.json exists for runs to be scored. The viewer can show ungraded outputs. Grading agents (agents/grader.md, analyzer.md, comparator.md) are instructions for an available agent, not installed subagents or an automatically running workflow.

5. Improve description triggering when justified

Separate task quality from trigger accuracy. Create realistic positive, negative, and borderline queries for the selected host; do not overfit to a single repository or a few phrases. The supplied assets/eval_review.html provides a local human-review template. For Claude Code, the preserved scripts can automate trigger evaluation and description proposals. For Codex, review actual host discovery/invocation traces or compare manual examples; do not pretend that text matching or explicit skill injection measures auto-triggering.

Output

Deliver the requested SKILL.md, directory/archive or edit to the authorized source. Summarize its trigger, target hosts, bundled helpers, and any limitations of execution. Do not auto-publish, install, push, merge, or create additional reports. If a requested destination is inaccessible, provide the complete source package and state exactly what was not changed.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.