Guide · Agent skill
How to use the Skill Creator agent skill (skill-creator)
Updated · written with AI from the project's own documentation and checked against our catalogue data
Skill Creator is an open-source agent skill, licensed MIT, from the bytedance/deer-flow project. It teaches an AI agent how to build other skills, which are reusable instruction folders that tell an agent how to handle a particular task. This guide explains what it does, how to install it, how to trigger it with plain requests, and what to check before running its scripts, so you can draft, test and refine your own skills.
Skill Creatorbytedance/deer-flow · MIT Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP from this page. Confirm it contains SKILL.md and the scripts folder.
- 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, upload it under Settings → Capabilities → Skills instead.
- 3Check requirements. Make sure Python 3 is available, Claude Code CLI is authenticated, and you have a local browser for the review page.
- 4Describe your goal. Tell the agent what skill you want, or share an existing draft. Answer its questions about inputs, outputs and success criteria.
- 5Test, review and iterate. Run the test prompts, open the generated review page, give feedback, and let the agent rewrite the skill. Repeat until you are satisfied.
What are agent skills, and what does Skill Creator make the agent do?
An agent skill is a folder of instructions that an AI agent can load when a task calls for it. Skill Creator is a skill whose job is to help the agent write those folders. Its own description says it can create new skills, modify and improve existing skills, and measure skill performance.
Once it is installed, the agent works through a loop: decide what the skill should do, write a draft, create a few test prompts and run the agent with the skill on them, evaluate the results, rewrite the skill based on feedback, and repeat. The source also describes expanding the test set and trying again at larger scale.
- Draft a new skill from scratch or from a workflow already in your conversation
- Write test prompts and quantitative evals to check the skill works
- Generate a browser-based review page to look at results
- Benchmark skill performance with variance analysis across runs
- Optimize a skill's description so it triggers at the right times
Install the Skill Creator skill
Download the ZIP file from this page. Then choose one of the install locations below. The skill folder should contain SKILL.md along with its agents, references, scripts and eval-viewer directories.
For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.
- Personal: ~/.claude/skills/
- Per project: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills, then upload
Requirements before you start
The skill's scripts need Python 3. The review server and benchmark aggregation use only the Python standard library, so no extra packages are named in the source.
The description improver calls the claude command, so you need Claude Code CLI authentication set up. You also need a local browser to view the review page that the skill generates.
- Python 3
- Claude Code CLI, authenticated
- A local browser for the generated review page
Trigger the skill with natural requests
You do not need special syntax. The skill's description lists the situations where it should load, so describe what you want in ordinary words. The agent then figures out where you are in the process and helps you move to the next stage.
If you already have a draft, say so. The agent can skip straight to testing and improving it. If you are starting from nothing, describe the task and the agent will ask questions about edge cases, input and output formats, and success criteria before writing test prompts.
- 'Help me create a skill from scratch that formats meeting notes.'
- 'I have a draft skill. Run evals on it and tell me where it fails.'
- 'Benchmark this skill across several runs and show the variance.'
- 'My skill never triggers. Optimize its description for better triggering accuracy.'
Review results and improve the skill
Skill Creator combines qualitative and quantitative checks. While test runs happen, it can draft evals if none exist, explain them to you, and use the eval-viewer/generate_review.py script to open the results for you to read alongside the metrics. Your feedback then drives the next rewrite.
The bundled agents grade outputs, compare two outputs without knowing which skill produced them, and analyze why a winner won. The grader also critiques the evals themselves, pointing out weak assertions that pass trivially.
- generate_review.py builds the review page
- aggregate_benchmark.py summarizes benchmark statistics across runs
- improve_description.py helps tune the trigger description
- generate_report.py is included for reporting
Safety notes: scripts and network access
Skill Creator runs scripts on your machine and uses the network, so treat it like any code you install. The bundled scripts execute code locally, and the source warns against running it on untrusted skill folders.
Evaluations can take time and consume model usage, because they run prompts through the agent. Only run them on skill folders you trust, and check the scripts before running them if you are unsure what they do.
- Install only from a source you trust
- Do not point it at untrusted skill folders
- Expect evaluations to take time and use model usage
What people use it for
- Packaging a repeated report-writing workflow into a skill that any future chat can load
- Checking whether a skill actually improves output by comparing runs with and without it
- Fixing a skill that never triggers by tuning its description
- Measuring variance in results across several benchmark runs before shipping a skill to a team
Frequently asked questions
Is Skill Creator free to use?
The skill is published under the MIT licence, as stated on this page, and the license file is included in the package as LICENSE.txt. Running evaluations uses model usage, so you should expect costs from the model you run with.
What do I need to run the skill-creator skill?
You need Python 3 for the scripts, Claude Code CLI authentication for the description improver, and a local browser to view the generated review page. The source does not list other requirements.
Where do I install the skill-creator skill?
Unzip the downloaded ZIP into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, upload it under Settings → Capabilities → Skills.
Is Skill Creator safe to run?
It runs scripts that execute code on your machine and it uses the network. The source advises against running it on untrusted skill folders, so review the scripts and only use sources you trust.
Why won't my skill trigger, and how do I fix it?
Triggering depends on the skill's description. Ask the agent to optimize the description, which uses the improve_description.py script. Note that this script calls the claude command, so Claude Code CLI must be authenticated first.
Is Skill Creator only for developers?
It is aimed at people who want to package a repeated workflow into a skill, teams that want evidence a skill helps, and people refining a skill's wording. If you only want to use existing skills, you do not need it.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Skill Creator suits developers and teams who want to build reusable agent skills and check with measurable evidence that they work. It is not needed if you only use existing skills, and you should avoid running it on skill folders you do not trust, since its scripts execute code on your machine.
Download the skill