A useful skill setup starts with a task you can finish and inspect. Start with one change in one repository, rather than installing a large collection and hoping the agent selects the right tools.
Define a finish line
Write down the input, the expected output and what you will check. For a settings screen, the input might be a design brief and an existing application. The output is a working form. Acceptance includes keyboard operation, a narrow mobile layout, and the application’s relevant tests. A screenshot can support a visual review; it does not prove that saving settings works.
Select by job, then read the source
The Agent Skills format packages instructions in a SKILL.md file and may include supporting resources. The client decides how it discovers and uses those instructions. The same folder format does not mean every client has the same tools or permissions.
Use the Agent starter collection to compare documented roles. A frontend instruction set can help produce the interface, while a testing skill can help plan checks. Open each detail page, follow its original source and inspect the version and prerequisites. Choose the smallest set that covers your finish line; add another only when a concrete gap appears.
Give the agent a bounded brief
The following is an illustrative request, not a recorded run:
Add a notification preference form to the existing app.
First inspect the app conventions and describe the proposed change.
Use the selected frontend skill only after reading its requirements.
Check keyboard use, validation errors and a 390px-wide layout.
Run the relevant existing checks. Record exact commands and results.
Separate observed results from assumptions and untested behavior.Ask the agent to inspect the repository before choosing commands. A generic test command copied from another project may be irrelevant. If the skill needs browser access, a plugin or a remote account, complete that setup separately and grant only what the task needs.
Review the result in layers
- Read the change: does it solve the requested task without unrelated edits?
- Exercise the behavior: do valid and invalid inputs work?
- Inspect the interface: can a keyboard and a narrow viewport use it?
- Record remaining gaps: an unavailable test environment is an unknown result, not a pass.
Keep the brief and the result together. If you repeatedly explain the same repository conventions, turn that procedure into a small team skill. Check the client’s current documentation for installation; SKILLPI’s collections are editorial guides and do not install software for you.
Continue with a release workflow
Once the change behaves correctly, the Product launch collection connects implementation to a release decision. Our release evidence guide shows why passing local checks and accepting a live release are different claims.