Writing

I counted the things I had to tell my agents twice

· 2 min read

I counted the things I had to tell my agents twice

The cheapest bug report on your agents is your own chat history. Count the messages you had to send twice and you have a list of what to fix.

I’m building games with Claude Code and Codex. Claude does most of the code and Codex checks it. I kept sending corrections I’d already sent, so I had an agent read my side of those chats and write them all down.

The same complaint, two sessions

September 12, in a Codex session:

why do you stop if you ltierally need to do all that UNTIL its done, everything, the WHOLE thing

The 25th, in another one:

why u quit, your task is to do the whole thing

Different sessions, same complaint, both from the game work. Agents stop partway through a list and I keep having to tell them to finish it.

I wanted the file in the starter

I spent four or five days preparing the studio and the workflow. I wanted an agent to go through all the Claude and Codex sessions and find the common caveats and bugs: where context got eaten, where something I said wasn’t clear, where the agent was being stupid. I wanted the result saved in the starter, so the next project begins from it.

The file landed today. It’s called PITFALLS.md, titled for what I had to say more than once, and it covers Claude and Codex threads from September 10 to 28. Each line has my words, a count of how many separate times I raised it, and an arrow to the doc that owns the fix. Only threads I led are counted, with no subagent packets. The busiest build days are sparse on disk, so the counts run low.

The robot walks away from a half-stacked tower of blocks

Stopping partway is the big one

The top section is about agent behaviour at every stage:

  • stopping partway through a task: x5
  • inventing work or jargon: x3
  • leading an audit with hints: x2
  • pushing on with a bloated context: x2

The audit one is from September 15. I wanted a Codex prompt that checks everything for bugs and optimizations, and I told it not to lead Codex with anything or hand it my context. The bloated context one is from the same day: I wanted a new agent because the current one was stupid with bloated context.

Some repeats became rules

A few of them are numbered rules now, in the starter’s RULES.md. Number 12:

Second identical message from Dee -> stop, name the broken rule, ask for /clear + one-line restatement.

The PITFALLS line that points to it says a point raised twice means the first fix missed. The same rule says a second fresh-context relaunch on the same subtask means re-scoping with me instead of relaunching.

The file only covers the game work

The same complaints show up in my other projects, and the file doesn’t reach them. On my media canvas app I asked on September 16 for a full check pass on the agent’s side before it calls something done.

The file is from today, so I can’t tell yet whether the rules fix anything.