HavenSkills

Claude Agent workflow skills

Four skills for people extending the agent rather than using it: building a skill, building a tool server for it, testing what you built, and getting straight answers about the API underneath. These are the ones that reward reading the source before you install, because they change how your agent behaves rather than just what it can open.

Skills
4
Full write-ups
1

One limit to know about

A skill only acts when its description matches what you asked for, so a vague description is the usual reason an installed skill never fires. skill-creator exists to close that loop: it writes test prompts and compares runs with and without the skill, instead of trusting the wording.

Filter

Recommended starting points

All skills in this category

skill-creator

Create a SKILL.md, write test prompts, and compare agent runs with and without the skill.

by Anthropic Skill guide
npx skills add anthropics/skills --skill skill-creator
Author
Anthropic
Repo
GitHub
Write-up
Read →

mcp-builder

Build an MCP server that gives an AI agent access to an external API or data source.

by Anthropic Skill guide
npx skills add anthropics/skills --skill mcp-builder
Author
Anthropic
Repo
GitHub
Write-up
Read →

webapp-testing

Test a running web app with Playwright. Click, fill forms, verify results, and capture screenshots.

by Anthropic Skill guide
npx skills add anthropics/skills --skill webapp-testing
Author
Anthropic
Repo
GitHub
Write-up
Read →

claude-api

Build with the Claude API using examples for messages, tool calls, and streaming.

by Anthropic Skill guide
npx skills add anthropics/skills --skill claude-api
Author
Anthropic
Repo
GitHub
Write-up
Read →

Instructions, tools and verification are three different problems

The four skills on this page look similar and answer different questions. skill-creator is about writing reusable instructions well. mcp-builder is about giving an agent a tool it does not have. webapp-testing is about proving that something runs. claude-api is about the API those instructions and tools end up calling. Installing all four because they share a repository is how people end up with a skill that never fires.

A useful way to read the page: if the agent does not know how to do something, write instructions. If the agent cannot reach something, build a tool. If you cannot tell whether either worked, test it.

What each one is for

Skill Question it answers Proof it worked
skill-creator How should this instruction be written and described? Two runs, with and without the skill, compared
mcp-builder How does the agent reach this API or data source? The tool is discovered and returns a real response
webapp-testing Does the workflow actually run in a browser? Steps, assertions and a captured screenshot
claude-api How does this API call work in practice? Working example with messages, tools and streaming

Where these go wrong

  1. A skill that never fires. A skill only acts when its description matches the request. Vague wording is the usual reason an installed skill does nothing, and it is also the easiest thing to fix: write test prompts and compare runs with and without the skill.
  2. An MCP server nobody can call. Auth, rate limits and timeouts get discovered at the first real call rather than in design. Test one tool with a real credential before building the second.
  3. Testing the development server. A local dev build and a production build behave differently in a browser. Test what you will ship, or at least state which one the screenshots came from.
  4. Reading the API from memory. Version, parameters and streaming behaviour change. claude-api is there to check the current shape instead of recalling it.
  5. Treating a green run as coverage. One success case proves the happy path. Add an empty or invalid input before you trust the workflow in regular use.

Install and verify in four steps

  1. Install the skill that matches the problem in front of you, not the whole set. Reload the agent afterwards so it can see it.
  2. For skill-creator, write three test prompts that should trigger the skill and three that should not, then compare the runs.
  3. For mcp-builder, call one tool end to end with a real credential and log the request and the response.
  4. For anything built with these, keep the fixture: a known input, an expected output and the date it passed. That is what makes a later failure diagnosable.

Choose a skill for your agent workflow

Decide whether you need to write reusable instructions, connect an external tool or verify a running application. Each outcome needs a different test.

How to check the result

Use a fixture with known inputs and expected results. Check one success case plus an empty or invalid input before putting the workflow into regular use.

Frequently asked questions

What is the difference between a skill and an MCP server?

A skill supplies reusable instructions for an agent. An MCP server exposes tools or data the agent can call. A workflow can use both.