The design kit behind Excellent’s admin, and behind screen-map, selfdocs and pressroom: tokens as CSS variables and plain components, light and dark, with a living styleguide. No framework and no dependencies, so a page a command-line tool serves looks the same as a React app.

A block for CLAUDE.md that stops cleanup agents from deleting model stores, keys and evidence, plus two disk lessons: measure free space with df, and a restart wipes /tmp.

A cleanup agent once deleted my local model caches in the middle of a live run.

Six rules, a block to paste into AGENTS.md, and a 60-line Claude Code hook that blocks git add -A, git commit -a and a bare git stash in a shared checkout.

Almost every multi-agent mess I’ve had came from three git commands that are safe for one person.

One command screenshots every route of your running app, draws how the screens connect and finds the pages nothing links to.

It finds the screens your navigation forgot, which a link crawler can’t see.

Every step in a plan ends with a “Done when” line and an “Evidence” line, so nobody, agent or person, can report a step done without showing it.

Checklists say what to do. They rarely say how anyone will know it was done.

Scans a repo and drafts its architecture docs with your local claude CLI, then serves them in a viewer. Your own notes survive every regeneration.

Regenerating is a merge: delete a section and your note sets itself aside, bring the section back and the note goes back in.

The writer, the editor model, the validator that checks every number and unit, and the frozen test week behind an AI-written newsletter.

The model never supplies a fact. It picks events by ID, and the template prints the numbers.

Write posts as MDX on your machine, preview them, and publish each one as a git commit that holds only that post. No dependencies.

pressroom publish <slug> commits exactly one file and pushes it.

A local, searchable archive of your X likes, with the text read out of liked images, the linked articles saved, and an MCP server so Claude can search it.

Most of what I like on X is screenshots of code and text. Now Claude can search them.

Synthetic users drawn from your customer profile try to do the job they came for in your product. You get a verdict, and every bug and dead end they hit, written up as work to fix. Today it only tests Excellent; making it work on any app comes first.

One shared, cached read of your Claude plan usage, so every status line and meter on your machine stops asking for it on its own. It reads an undocumented endpoint, so it may break.

However many status lines are open, it’s one request per interval.

Walks an agent from the objects and workflows of an app to its structure, page grammar, screen inventory and design system, then a hard critique. Each stage says what done looks like.

A read-only review that asks one question of a change: does it make the codebase simpler or messier? It flags 1,000-line files, thin wrappers and tangled branching.

Rules that fail a post for em dashes, “X, not Y”, hype words, emoji and any number that isn’t in its source.

My own posts have a linter. “X, not Y” fails the build.

Running a personal AI assistant on a $25 a month server: Tailscale, Docker, Claude, a calendar and text messages, with a table for locking it down.

A spaced-repetition trainer for Mac keyboard shortcuts: switching apps and windows, the terminal and the system.

Also the story of why a web page can never catch Cmd-Tab.

A one-file tracker for a founder’s posts: drafts, a calendar, the numbers and a report on what works.

I built my own instead of paying for one. I’m using it for a month before I share it.