I can't read what my agents write, so I changed the instructions file

I have ADHD and I can’t read the walls of text coding agents write. So I stopped fixing it chat by chat and put the fix in the instructions file every agent reads when a session starts.

For months I’ve typed the same thing into whatever chat was open. “tldr thats too long” is the whole fix I get to send. It works for one reply. The next session starts from zero and the wall is back.

One chat at a time doesn’t hold

I have ADHD, so I want less cognitive load, and I ask for short answers everywhere. In May it was a project registry that had to be easy for the agent and easy for me. In June, in a media project, it was a tldr of what we were building. In July I asked one agent what to tell another so I’d get simpler questions and less to read.

August was the worst stretch. In an app project I called one reply too much to read and asked for it concise. In another I wanted the next steps one at a time, short, because I’m ADHD. Slide copy came back as a wall of text for an audience that doesn’t read much. Then “tldr thats too long”.

Every one of those fixed the chat I typed it in and nothing else.

The old rule was one soft line

Claude Code had a single line telling it to keep turns concise and not narrate every thought. Codex, Kimi and Cursor had nothing. I still got walls of text.

Kimi is another coding agent I use, and it reads its own AGENTS.md like Codex does. While I was reviewing my setup with the mentor skill, the agent read the Claude docs and reported that thinking effort doesn’t set reply length. Length has to be asked for in the instructions. That fits what I told it: I’m fine with long thinking, I just can’t read long output.

feel free to think when thinking but my end result should be clear for me so i dont read a lot

An Output block in four files

I put an Output block into four files: CLAUDE.md for Claude Code, an AGENTS.md each for Codex and Kimi, and the Cursor rules file. The old soft line is gone.

The first version was three short sentences. Lead with the result. Bullets or tables over paragraphs. Bold the one thing to act on. Then the agent rewrote it in all four files into a version with an example reply, and I kept that one:

Done: X. Changed a.py:12, b.py:40. **Next:** run tests.

The shape is five bullets. Replies are focused and brief, with the answer and not the caveats. The result comes first, then three to seven bullets or a table, then one bold next action. Explanations stay high-level unless I ask for depth. Replies point to file:line instead of pasting code or logs. The agent wrote the shape, not me.

Each agent applies it to itself

I want Codex and Kimi to apply the rule themselves, since each knows better how to use its own instructions file. A few hours later today I asked for a short summary of everything they’d need, one I could hand to Codex so it updates itself, and then the same for Kimi. Then I had both checked for bloat and polished.

The same session also set up a statusline. I wanted to see only what I need to see.

It’s in place today. I’ll see if the walls of text stop.