What you’ll make
Make a small Claude skill for one task and prove it handles a missing detail.
Worked example
A skill that drafts a changelog entry
A planner app developer provides a release identifier and a list of reviewed changes. The skill returns a heading, a user benefit, and the exact menu label if it is verified. The supporting example shows the shape of a previous entry.
The complete request produces a draft for the release. A case with an unapproved change asks for review status. 'Sort these export bug reports by severity' stays a triage request because no changelog was requested.
Before you start
- Pick a process you already run manually and write down two good outputs, one incomplete case, and one task that uses similar words but should follow a different workflow.
- Choose where you will run the skill. For Claude on the web, confirm that custom skills and code execution and file creation are available in your account. For Claude Code, use a project folder you can edit.
Choose one repeat job
Write a request you’d actually use and what you want back. List the facts it needs, plus one similar request that belongs to a different job.
I ask for a changelog entry from reviewed release changes. Sorting export bugs belongs in triage.
The skill’s job is clear enough to recognize from one request.
Draft and save the skill file
Copy the prompt into Claude. Attach a short approved example if you have one. Read the draft and save the file as SKILL.md in a folder named for the skill.
I supply my release number and an example heading and benefit.
The file should say how to write that entry and when to ask if a change was approved.
Draft one small Claude skill
Write one instruction-only Claude skill. Use the details I provide in this chat or in my attachments. Return the complete SKILL.md, with name and description frontmatter, the requests it should handle, the inputs it needs, and the steps for returning the result. Use current request facts and approved sources. Give a specific question to ask if a required input is missing. Keep the task narrow and leave nearby requests to their own workflow. Ask one focused question before drafting if the trigger or result is unclear. Omit scripts and return the file content in one code block.
Make the skill available for a test
In Claude Code, put the skill folder under .claude/skills/ in your project. If you use Claude on the web and your account supports custom skills, ZIP the skill folder, upload it under Customize and Skills, and enable it. Open a new chat to try the sample request.
Try a complete request and one gap
Run the sample once with the required facts and once in a new chat with a key fact removed. Compare what comes back with the rule you saved.
- The complete request gives you the expected result.
- The incomplete request asks for the missing fact.
- You kept the skill file and both requests for a future test.
Optional: Try a similar request that belongs elsewhere
In a fresh session, give Claude a request that sounds like the skill’s job but needs a different result. Tighten the description if the skill gets picked for both jobs.
Write a skill for the job I keep doing
Give Claude a repeat request, its must-have inputs, and an approved example to draft your SKILL.md.
I want an instruction-only skill for [repeat job and deliverable] in [Claude on the web or Claude Code]. It should apply when someone asks [real request examples] and leave [similar request with a different goal] to another workflow. Each run needs [required inputs], can use [optional inputs], and should follow [source priority]. Here is an approved sample [file]. Use it to learn the structure and tone. Take names, dates, and facts from the current run. Return a proposed folder tree and complete SKILL.md with YAML name and description, when to use it, how to ask for missing facts, the procedure, the output shape, and observable final checks. Reference a supporting file by path only if it has a specific job. Keep the draft in chat and point out anything I need to decide before installing it.
Give me three requests that could prove it works
Use the draft skill to make a test sheet before trying it in separate fresh chats.
From [SKILL.md, example, and target environment], write three paste-ready requests for fresh chats. Include a complete case, the same kind of case missing exactly one required fact, and an adjacent request that should use another workflow. Give me expected output or exact behavior and an observable pass condition for each. Use sample data I can verify. Include a direct-invocation check if my environment supports it so I can tell a trigger issue from a procedure issue. Return a results table I can fill in and tell me which single section to inspect first for each likely failure.
Troubleshooting
Why is the uploaded skill rejected?
Inspect the ZIP contents and confirm the skill folder contains the instruction file with both required frontmatter fields. Rebuild the ZIP with the folder at its root.
Why does it show up on every writing request?
Tighten the description to include the distinct input and intended output. Add a near-miss example to the body, then repeat the unrelated test in a new chat.