Skip to content
Commerce Fieldbook
Fieldbook/Start here

How to use agent skills

Treat a skill as a written procedure with known inputs, checks, and limits.

stableReviewed 2026-07-25

An agent skill packages instructions, references, and optional scripts for one defined job. The Fieldbook also publishes SKILL.md files for tools that support that format.

Before you run one

Make sure you are allowed to inspect the page, system, dataset, or customer journey. Payment tests and destructive actions belong in approved test environments.

Gather the listed inputs. Missing facts should stay missing; the agent should not fill them with guesses.

While it runs

Ask for the source and timestamp behind each material finding. Keep observations apart from suspected causes.

Customer-facing edits, campaign changes, and production writes require explicit approval. A diagnostic task should not quietly become a publishing task.

Review the result

Check that the response follows the requested format, cites the source material, and leaves unresolved questions visible.

Then rerun the relevant check against the same product, page, market, or journey. A fix is not complete until the original failure path has been tested again.

Model portability

The task description is independent of one model. Claude-compatible metadata may appear in the download, but the procedure can also be used by coding agents, analysts, QA teams, or a person following the checklist.