Bring the tool to where people work: shipping an agent skill

MarketingPatcher started as a website. You paste a URL, you get a report, you go fix things. The last step was always the weak one.

The gap between the report and the fix

Most people who run an audit are building their own site. They read the report in one tab, then switch to their editor and try to work out where the meta description lives in their framework. Some fixes are one line. Others need a new file, a config change and a redeploy. A good share of reports never turned into changes at all.

Meanwhile, more and more of those same people were writing code with an agent like Claude Code, Cursor or Codex. The agent already knows the codebase. It just did not know what to fix.

Shipping the audit as a skill

So we packaged the audit as an agent skill. It is a small folder with instructions and two scripts. The agent runs the same audit the website runs, reads the report, and patches the failing checks in your repo.

If the site is not deployed yet, the skill falls back to checking the code itself, using the same rules. The scores match the website exactly, because both use one set of checks.

What we learned

Keep the instructions short and concrete. Agents follow a numbered workflow well and vague advice badly.

Treat everything fetched from the web as data, never as instructions. The skill tells the agent this explicitly, because an audited page could contain text written to steer it.

Respect rate limits in the instructions, not just the API. Telling the agent never to retry a 429 saved more quota than any server-side rule.

Try it

Install it from the skill page and ask your agent to audit your site. It works anywhere skills do, and the web audit is still there when you just want the report.