Save the method, including the checks
A skill gives the agent instructions for a recognisable job. I keep the starting material, the work to do and the evidence needed at the end together. The valuable part is the review standard: a repeated task should not need a new definition of what counts as finished.
The names are specific because the inputs differ. carbon-copy works from a live site, but image-parity works from a static reference. I choose between them by what I have to inspect, so the agent gets the right source and the right way to compare its work.
For an agency, that source might be a working website or a client’s flat artwork. Keep the reference and the checks attached to the job. A reusable instruction needs to say what it starts with, so the agent can ask for missing material before it starts building.
Choose the skill by the job
- tech-lead — plan, brief, verify, integrate and ship product work.
- carbon-copy — measure a live site and rebuild it in React.
- image-parity — match a component to a static reference.
- design-system — extract tokens, components and contracts from a codebase.
- lead-magnet-pdf — research branded PDFs; the renderer “fails on overflow”.
- ecommerce-operator-articles — research ecommerce articles with a source ledger and adversarial fact check.
withseismic-grid gives a WithSeismic page a layout the agent can check by measurement. vnaut-redesign takes a part of Vouchernaut, inspects it at four widths, names what needs fixing, rebuilds it and verifies it. Each carries a standard I would otherwise have to explain again.
A worked example: tech-lead
I give tech-lead a product job and the repo containing it. It reads docs/DECISIONS.md for settled choices, docs/BACKLOG.md for the queue and docs/prds/ for the work itself. The current brief supplies the goal, the paths to read, the boundary, dependencies and checks.
It chooses the next job, writes or updates its PRD and builds in the current session unless delegation clearly pays. Acceptance criteria use observable WHEN/SHALL behaviour. On withseismic.com, that means the lead can return to a specific agreement when it reviews a page or an interaction.
It reads the diff, reruns the checks and inspects the evidence. Each finished PRD is deployed in dependency order and checked in production with real data. The lead records the production result in Implementation Notes and updates the backlog, so another session can continue from the files.
The skill stops for spending, publishing or announcing, destructive changes outside the repo, product or legal judgement, or missing credentials. This site has my recorded permission to deploy and verify its PRDs. A skill supplies the working method; my permission sets what it may carry out.
Keep the stopping point in the skill
vnaut-redesign stops after verification until told to ship. That stopping point travels with the design instructions, so asking for a redesign has a clear end. I can review the result before deciding whether it should reach the live site.
The checks need to suit the output too. lead-magnet-pdf combines adversarial fact checks with a Chromium renderer that fails on overflow. The facts and the page layout both have to pass. Saving that requirement makes it part of every PDF job rather than something to remember at the end.
Start with a job you can judge
Start with a job you already understand well enough to judge. Keep its inputs, method, checks and boundary in the skill, but leave the current business decision in the brief. That is how I reuse the work without asking an old instruction to decide what matters today.
I keep 21 skills of my own in ~/.claude/skills, counted on 18 October 2026, with 23 skills on the Codex side. I use the name to find the method again and the current brief to supply the work. That lets me repeat a useful process while keeping the business decision with me.