Engineering2 min read

How to Know Your Harness

Same laptop, same Claude, two repos, two different agents. The difference is the harness.

·
aiclaude-codeagentstooling

Same laptop. Same Claude. In one repo, it checks with other agents before it edits a file. In the other, whether it mocked the database depended on the folder.

The model did not change. The harness did.

What a harness is

Everything the agent is given before you type: instructions, skills, memory, hooks, tools.

Nobody writes it in one place. It is stacked from three layers: your global files, the plugins you installed, and the repo you are in.

A harness is everything Claude is given before you type. It is stacked from three layers.

Project
  1. 1. Global / yours, in every project

  2. 2. Plugins / installed once, in every project

  3. 3. Project / teamroom, only in this repo

AGENTS.md

Instructions for this repo, in the file Codex and most other agents read.

Loads
Every session in this repo.
Lives in
teamroom/AGENTS.md
Change it
Write it at the repo root and commit it. To be sure Claude gets the same rules, add a CLAUDE.md with one line: @AGENTS.md

The shipping workflow used to sit in a Claude memory, where Codex never saw it. Now it is here.

What Claude gets in teamroom

  • Instructions2 (1 here)
  • Skills92
  • Memory0
  • Hooks8 (3 here)
  • Tools4 (1 here)
  • 92 skill descriptions: Claude may not see them all
Click any piece to see where it lives and how to change it. Faded bars are shared, solid is added by the project.

When each piece loads

Instructions load every session. Skills show only their description until a request matches. Memory loads its index, and a note when it seems relevant. Hooks run on events, whatever the model decides. Tools wait until called.

What loads every session costs attention every session.

How to read your own

In one folder, Claude Code can tell you. /memory for instructions and notes. /context for what fills the window. /plugin, /mcp and /hooks for the rest.

Across every folder, nothing could. So I built Klosti. Here is why.

What I found in mine

A blog rule saved in teamroom's memory, where the blog never sees it. A memory repeating a rule AGENTS.md already had. A type check that ran on new files and skipped every edit. One MCP server defined twice, and the hidden copy won. 112 skill and command descriptions in one repo, more than Claude may keep in view.

Most fixes were a move, not a rewrite.

The rule was right. The layer was wrong.

New posts, straight to your inbox.