Learn from the best AI operators you know

I run my seo-preflight skill on the staging link. It goes in this order and stops at the first failure:

1. Checks the page’s main term from keywords.csv is in the title, the H1 and the URL.
2. Loads the page with the Playwright MCP server and checks there’s no noindex, the canonical points to itself, and the meta description is under 155 characters.
3. Checks the schema markup has the fields Google asks for, like datePublished on a blog post.
4. Asks the Search Console MCP server which of our pages already rank for that term, and lists two to link from.

The day after launch I run URL Inspection in Search Console to confirm Google indexed it.

I start from what customers said, not from the feature list. Through the Intercom MCP server I pull the last 90 days of support chats about the problem we’re launching for, and keep five lines people wrote in their own words.

My launch-voice skill writes three headlines, each built on one of those quotes, then checks every line against banned-phrases.md: “seamless”, “unlock”, “all-in-one” and about forty more.

Last is the competitor test. I put each line on the homepages of our two closest rivals. If it would fit there too, I cut it.

Every line goes to the team with its quote beside it, so anyone can see where it came from.

I paste the frame link and Codex reads it through Figma’s MCP server. The order matters:

1. get_metadata first, to list every layer without pulling the whole file.
2. For each component it checks Code Connect, which maps Figma components to ours. If there’s a match it uses ours; if not, it stops and lists it for me instead of making a new one.
3. Colours and spacing come from get_variable_defs, mapped to our tokens. A lint rule fails any raw hex value.
4. It screenshots the build with Playwright at 1440 and 390 wide, next to the Figma export, and I compare them side by side.

Four: the two biggest in our space and the two newest. I don’t check them by hand any more.

A Claude Code routine runs every Monday at 8am. It scrapes each one’s pricing page, changelog and sitemap with the Firecrawl MCP server, with change tracking on, so it only gets back what’s different since last week. New URLs in a sitemap are usually landing pages they’re testing, so I read those first.

It adds a dated entry to competitors.md: what changed, the link, and one line on whether it affects us. A price change also opens a Linear issue for me. Quiet weeks just say “no change”, so I know it ran.

Four passes, and the agent that reviews isn’t the one that wrote the code.

1. pnpm typecheck and pnpm test. I read the output, not just the last line.
2. /review against main in a fresh Codex session, told to report bugs only and skip style.
3. With the Playwright MCP server I sign up as a brand-new user, then open each changed screen with no data and with the API returning errors. Each empty and error state gets a screenshot in the PR.
4. The release note is written from the diff in plain words, and I check each line against a commit.

Nothing merges until all four are in the PR description.