Capture a recurring workflow
Turn an established process with known inputs and deliverables into instructions an agent can follow in future tasks.
Skill Creator by Anthropic
Use Skill Creator by Anthropic to turn a recurring task into a reusable Skill and improve it through examples. Define when the Skill should activate, draft its instructions, compare test runs, and use feedback to refine the workflow and its description.
Install and get started
Choose your agent to see how to get started. In Floatboat, copy a request that asks the agent to read the original instructions. Other options show the Skill installation command or the publisher’s native setup steps.
Choose an agent
Start in Floatboat
Copy the request into a Floatboat conversation
Read the original Skill Creator instructions at https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md. Explain its purpose, required tools, and installation options, then help me use it.Let the agent read the original instructions and check how to use them in your environment.
Download Floatboat ↗Try this request
Help me turn our weekly customer-feedback summary into a Skill. Define the inputs and output, draft the instructions, and propose test requests before running an evaluation.
About Skill Creator
Skill Creator is Anthropic’s workflow for authoring and evaluating Agent Skills. It starts by clarifying the task, activation conditions, inputs, and expected output, then moves through a draft and representative test prompts. The source includes scripts for inspecting evaluation results and comparing runs, with repeated revision as feedback identifies gaps. This makes it useful for capturing a recurring work process or improving a Skill that activates inconsistently. Its evaluation workflow is written for Claude-based execution and needs the referenced tools to be available. This page describes Anthropic’s implementation; OpenAI and other publishers also use the Skill Creator name for different packages.
Use Cases for Skill Creator
Explore Skill authoring tasks with a clear recurring workflow or an existing draft to improve. Capture a repeatable process, evaluate whether instructions produce the intended result, or refine a description that causes the agent to miss relevant requests or activate at the wrong time.
Improve an inconsistent Skill
Compare representative runs to identify missing steps, unclear expectations, or examples that do not cover the task.
Refine activation behavior
Check whether the description matches realistic user requests, separating tasks that should trigger the Skill from nearby tasks that should not.
How to Use Skill Creator
Explain the recurring task and supply examples of successful work before drafting. Define the trigger and expected output, test the draft with representative requests in the supported environment, and revise the instructions using both user feedback and the results of those runs.
- 01
Define the task and examples
Describe the process, inputs, and expected output. Provide successful examples and any requirements the draft must preserve.
- 02
Draft and evaluate
Create the instruction and representative test prompts. Use the source’s supported evaluation tools to inspect how the draft performs.
- 03
Revise from observed results
Compare feedback and test outputs, then change the unclear parts. Rerun relevant cases and check that the description still fits the workflow.
Explore More Agent Skills
Read another Skill guide and check its original source before using it.
Skill Creator FAQ
Is this the OpenAI Skill Creator?
No. It is the skill-creator package from anthropics/skills. OpenAI’s package has a different source and workflow, even though both help users author Skills.
Do I need an existing Skill to start?
No. You can begin with a recurring task and examples. If a draft already exists, provide it and describe the behavior you want to improve.
Does the workflow run evaluations automatically in every agent?
The source’s evaluation steps use Claude-based runs and bundled scripts. Other hosts may read the instructions, but automatic evaluation requires checking their execution and delegation capabilities separately.
What should a test request include?
Use a realistic user request with the required input and an observable expected result. Include ordinary cases and important boundaries rather than only examples that repeat the draft’s wording.
Can it improve when a Skill activates?
Yes. The source includes a description-improvement workflow. Evaluate both relevant requests and nearby non-matching requests so better recall does not simply create more false activations.
What if a successful result is subjective?
State the audience and quality criteria, then review outputs with the user. Quantitative checks can support the review when they measure something meaningful, but they do not replace the editorial or design judgment.