How to Write an Agent Skill
Turn a recurring task into an agent skill with clear triggers, required inputs, review criteria, and two examples to test.
Write an agent skill after you have completed one real instance of the work. Keep the original input, an approved output, and a record of a case that had to stop. A second person should be able to start with the same material and know what a successful output looks like.
Choose a repeatable task
A product page brief is a useful example. Every new SKU arrives with some mix of photos, specs, prices, and delivery terms. The reusable work is checking that information, deciding what belongs on the page, and preparing a brief a designer can use. Separate values that change every run, such as a sale price, from rules that stay true, such as recording the source and expiration of a sale price.
Describe the trigger, input, and output
The open Agent Skills format uses a SKILL.md file with name and description metadata. Put the phrases people actually use for the work in the description. In the body, state the minimum input, order of operations, output location, and acceptance conditions. For the product brief, an incomplete specification might produce a short list of missing fields and a draft with those fields flagged. Keep long examples, field definitions, and reference material in supporting files. Agent Skills specification
Automate checks that have a definite answer
A script can confirm whether files exist, image names collide, or required columns are present. Leave judgment calls with the reviewer, including whether a photo shows the correct SKU and whether the proposed claim matches the evidence. SCTCamp’s public landing page project shows how source assets, page sections, and a visual review fit together. The talking-head editing project provides a second example with a different input and output. SCTCamp landing page skill · Talking-head video factory
Test a successful run and a blocked run
Run a complete input through a new conversation or ask a teammate to run it. Check that the output appears at the expected path and can go straight to the next step. Next, remove a required fact or the source for an image. The skill should explain what is missing and stop before publication. Keep both examples near the skill and rerun them after changing the procedure. Document who can approve an external action and which evidence they need to see. Your first AI commerce workflow · SCTCamp course