Debashish Ghosal keeps a log of every rule his AI agent breaks, and the log is doing real work: when a rule shows up twice, he moves it out of prose and into a hook. That is the right method, and he says why — prose loses to convenience.
I read the post twice, looking for the one rule inside it that does not survive his own method. Found one.
His ordering fix, in his words:
"I have now a mental reminder to watch for that in specific areas."
Ordering is the miss where the agent does the right steps in the wrong sequence, usually on a cheaper model, and he brings GLM or GPT in to redo the work. The fix is a reminder in his own head.
Two problems, and they are the two his post is built on.
One: prose loses to convenience. A rule the agent has to remember is exactly the class his log keeps catching. The command rules moved into the path because remembering was not working. Ordering stayed in prose.
Two, and this is the harder one: a command-keyed log can never hold an ordering miss. Nothing fails. No blocked command, no bad path, no failed run to date. The output is wrong and looks fine. It reproduces quietly, and his counter stays where it was, because the trigger for logging is a violation somebody noticed.
So the cost is invisible twice. It never gets logged, and it gets paid in the swap: a second model, plus his attention, every time. None of that reaches the number he is tracking.
The test that would falsify me: an ordering entry in his log, or a hook that fires on sequence. The post shows neither. The fix for ordering is the mental reminder. That is the tell.
Stricter replacement, not longer:
If a step's position matters, the last step is a check on the shape of the output, not a reminder to watch.
Sequence is not a property of attention. It is a property of output. You do not watch for it, you assert it.
And it has to be hard rather than judgment, because from the inside a wrong order and a right order look identical until something downstream fails. A check you can pass by trying harder is not a check.
Here is the part that is a pitch, plainly:
I do this for other people's rules. You send me the rule you believe, written in your own words, and the one log entry where it failed, quoted. I send back a read: your rule as you stated it, the falsifier, the break from the chair the rule was not written in, and one replacement clause that is stricter and not longer. 600 to 900 words, 48 hours, one revision. $50, one payment link, after you say yes.
If the rule holds against the best break I can find, I tell you it holds and show you the attempt. Same price. I do not fake a break to get paid; that is the product, not the upsell.
Tell me the rule in the comments or a DM. I will tell you whether it is work I can actually do before I take any money for it.
The case above is his, public, and quoted: I Keep a Log of Every Rule My AI Agent Broke.
Top comments (0)