Harness Starter — AI skill for Claude Code

Harness Starter

This is a working harness. Not a worksheet, an actual folder you can open in an agent and ask questions of right now.

The architecture is tool-agnostic:

• one identity file
• one map file
• one real domain with actual data inside

The only thing that changes across tools is the instruction filename. The system itself stays the same.

What's here

• me.md — who the person is, what they are building, and how they work.
• content/ — a real example domain with ideas, scripts, and performance notes.
• logs/ — weekly working notes.
• CLAUDE.md — the Claude Code version of the map.
• AGENTS.md — the Codex-style version of the same map.

Which file does my tool read?

• Use CLAUDE.md for Claude Code.
• Use AGENTS.md for Codex and other AGENTS-style coding agents.

If your tool expects a different instruction filename, keep the same map and copy it into the file that tool reads.

Try it before you change anything

1. Open this folder in your coding agent.
2. Ask: what should I post this week based on what's been working?
3. Watch it read the map, find that content/ and logs/ are relevant, and answer from the actual files.

Nothing is stored. Everything is re-read from the folder in that moment.

That is the harness. The answer is good because the files are good and the map told the agent where to look.

Folder structure

.
├── AGENTS.md
├── CLAUDE.md
├── content
│   ├── ideas.md
│   ├── performance.md
│   ├── README.md
│   └── scripts
│       └── video-3-how-to-build-it.md
├── logs
│   └── week.md
└── me.md

Make it yours

1. Rewrite me.md so it is actually you.
2. Replace the example files in content/ with your real domain.
3. Keep one live map file for the agent you use, or keep both map files in sync if you use multiple tools.

Keep it to one domain at first. One folder, one domain, one map. Add more only after this already feels useful.