@lukasen_xyz: pstack's new /correct command is designed to fix one specific issue: the same mistake you tell the agent about multiple…

pstack’s new /correct command is designed to fix one specific issue: the same mistake you tell the agent about multiple times, and it goes off and makes it again somewhere else next time.

Its approach isn’t to keep correcting it with you in the conversation, but to go back through the commit history, identify the pitfalls the agent keeps stumbling into, and then modify the codebase itself: tweak the architecture, add type annotations, insert checks. If the same type of error crops up again, the checks will flag it directly—no relying on the agent’s memory.

poteto added a line today that I really agree with: “Before there was agent-generated crap code, there was human-written crap code.” Backend devs getting dragged into writing frontend stuff and producing crap code is totally normal; big companies have been dealing with this forever, relying on linting, compiler diagnostics, tests, and those kinds of constraints. Now that agents are here, it’s everyone’s problem. At the end of the day, it’s just good engineering practices—nothing new.

This really resonates with me. Adding line after line to CLAUDE.md like “don’t do this” is basically just begging it to behave; turning it into a lint rule or a test means it can’t dodge that class of error anymore.

You don’t need pstack to follow this approach. Here’s a snippet I wrote myself that you can just feed straight to Claude Code or Codex:

“Go back through the recent commits and our conversation, identify the classes of errors I’ve had to correct you on repeatedly. For each class, propose one way to automatically block it: a lint rule, type constraint, test, or check script—don’t just add it to the prompt. First give me a plan, noting how many times each class of error has occurred, and don’t jump straight to changing code.”

pstack link posted in the comments.

lauren (@poteto): everything i know about managing agents i learned from the amazing programmers and computer scientists that came before

constraints in your codebase are freeing for both humans and agents! this has always been true, but if you were a small dev team you likely never had to think