Step 01
Open Terminal and Start Claude Code
Open a terminal in the trusted skills repository and start Claude Code there. Confirm the current path, visible skills, pending Git changes, and permissions before asking it to create anything.
IntermediateEngineering
How to Generate and Use Custom Claude Skills via the Command Line with Claude Code
Use Claude Code inside a local skills repository to generate a task-specific skill, inspect and validate the files it creates, then test the skill on representative input and improve the reusable instructions.
Try this with your AI assistant
Download the skill for this workflow. Version 1.0.1.
Source episode · 18:37Claude Skills Explained: My Workflow for Creating Reusable AI Agents with Cursor and Claude Code
with Claire Vo
Claire starts Claude Code inside her local skills folder, invokes the create-skill skill to build a changelog-to-newsletter skill, runs its validator, and tests it with an actual changelog.
Before you start
What you need
- Claude Code installed and authenticated
- A trusted local skills repository
- A working create-skill meta-skill
- A narrow recurring task with representative input
- An expected output format and examples of unacceptable output
What you’ll make
A task-specific Claude skill that Claude Code can discover, validate, and apply consistently to representative inputs.
Tools used
- ClaudeVisitClaude
Anthropic AI assistant
- Claude CodeVisitClaude Code
Claude for coding tasks
The workflow
Follow the sequence once, then adapt the prompts, checks, and handoffs to your own setup.
4 steps
Step 02
Prompt Claude Code to Use Your 'Meta-Skill'
Invoke the create-skill skill with one recurring task. Supply the real input shape, output format, tone or policy constraints, and examples that show when the skill should and should not run.
Example prompt
Use the create-skill skill to create [skill-name] for [recurring task]. Inputs look like: [representative input]. Produce: [exact output]. Follow: [constraints]. Trigger on: [matching examples]. Do not trigger on: [non-matching examples].
Step 03
Observe the Automated Generation
Review the proposed folder and diff before running scripts. Check the metadata, instructions, referenced files, trigger description, output contract, and read-only validation behavior.
Example prompt
Show the generated file tree and summarize the purpose of each file. Run only the read-only validation checks defined in the trusted create-skill folder. Report warnings and do not modify unrelated files.
Step 04
Test the New Skill Immediately
Test the new skill with representative input in a fresh turn, then try an ambiguous input and an unrelated request. Move recurring corrections into the skill and keep the improved example with it.
Example prompt
Here is the latest [input type]: [input] Apply the appropriate skill. Preserve source facts, follow the required format, and label missing information instead of inventing it.
If the output isn't perfect, you can easily open the `SKILL.md` file for the new skill in a text editor and tweak the instructions to refine the tone, format, or logic.
What good looks like
- The generated folder contains only files the skill uses.
- Claude discovers the skill for matching requests and ignores it for unrelated requests.
- Validation is read-only and the generated output follows the required format.
- Repeated corrections are made in the skill instructions and covered by a test example.
Build your next product with ChatPRD
Turn an idea into a PRD, user stories, and a plan.
Runbook notes
How to recover when the loop fails and where human judgment helps.
If it goes sideways
- Claude Code cannot discover the meta-skill or the newly generated skill
- Start it from the intended repository, confirm the supported skill path and metadata, and ask Claude to list the skills it can see.
- A broad brief produces a generic skill that triggers on unrelated work
- Narrow the task, add positive and negative trigger examples, specify the input shape, and define one expected output.
- The validator passes structure but the skill still produces poor work
- Add representative and adversarial examples, evaluate the actual output, and treat structural validation as one check rather than proof of quality.
- A generated skill or script reads credentials or unrelated repository files
- Inspect file access and scripts before running them, use synthetic inputs, constrain paths, and remove network or secret access that the task does not require.
Related workflows
intermediateDataMeta Muse
How to Use AI for Hobby Improvement and Data Analysis for Golf & Chess
Connect data from your hobbies (like golf scores or chess games) to an AI agent to get automated analysis, visual dashboards, and actionable coaching insights. Turn raw data into a story that helps you improve.
Open
beginnerPersonalMeta Muse
How to Create a Personalized Shopping Catalog from Your Instagram Saves Using AI
Turn your unstructured Instagram 'Saved' collection into a fun, personalized shopping catalog. This workflow uses an AI agent to organize your saves and present them in a nostalgic, engaging format like the SkyMall magazine.
Open
advancedContentMeta Muse
How to Revive an Old Blog or Digital Archive as a Personal AI Podcast
Resurrect your old digital content, like a forgotten blog from Blogger or Tumblr, by using an AI agent to find, consolidate, and transform it into a daily personal podcast. Reconnect with your past self in a new, listenable format.
Open
Start shipping better products.
Join 250,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.
Free to start · 14-day Pro trial · SOC 2 Type II
