My AI Skipped the Rules This Morning.
I Fixed the Instruction, Not the AI.
This morning one of our AI agents was handed a task and, within seconds, declared the file it needed "didn't exist." It did. The agent just hadn't synced to the latest first — and our rule that says sync first sat there, ignored. The easy move was to blame the AI. I did the opposite.
What actually happened
We run AI agents on real work now — not demos. One of them was handed a build task: "build this from the spec." It went straight to looking for the spec, didn't find it in its local copy, and reported back that the spec wasn't there.
The spec was absolutely there. It had been written and saved an hour earlier. The agent was just working from an out-of-date copy and never synced to the latest before it started. We even have a standard for exactly this — a start-of-session checklist whose first step is "pull the latest." The agent skipped it and jumped straight to execution.
The easy answer is to blame the AI
And it's the wrong one. Blaming the tool feels good and teaches you nothing. Worse, it's not even accurate. The agent did precisely what a rushed junior does on day one: it heard "build this" and ran at the task, skipping the boring setup step that didn't feel like the job.
If a new hire did that, you wouldn't conclude "this person is broken." You'd look at how you briefed them. The AI deserves the same honesty — especially from a company that builds with AI for a living.
The real cause: the instruction wasn't self-sufficient
Here's what I realised. We had the rule. But the rule lived in a long setup document the agent was supposed to read and remember. Meanwhile, the actual instruction I gave was four words: "build from the spec." That short instruction created a strong pull to start executing, and the one load-bearing step — sync first — was sitting in a separate checklist, competing with the goal and losing.
The instruction and the ritual were two different things, and when they disagreed, the task won. That's not an AI flaw. That's an instruction-design flaw — and it was mine.
The fix: make every instruction stand on its own
So I stopped relying on the agent to remember anything. I rewrote the brief itself to be self-sufficient — complete from start to finish, assuming zero memory and an out-of-date starting point:
- The first line is now "sync to the latest before you do anything — if a file looks missing, you haven't synced."
- The last line is the definition of done and what to report back.
- The short pickup instruction only points to the brief. The brief carries everything — including "sync first."
I didn't make the rule stricter. I moved the one step that mattered out of a skippable ritual and into the instruction itself, where it can't be missed.
This isn't really about AI
It's the oldest management lesson, just sped up. If your team keeps "skipping the rule," the rule is badly placed — not the team badly behaved. The instruction is the manager's responsibility, not the worker's failing.
AI just makes this brutally fast and visible. It does exactly what you wrote, instantly, at scale — so a vague or incomplete instruction doesn't quietly get patched by someone's common sense. It fails out loud, immediately. That's a gift, if you're willing to read it as feedback on your instructions instead of a complaint about the tool.
The bottom line
If you're putting AI to work in your business and it keeps doing the wrong thing, resist the urge to debug the AI. Debug the instruction first. Nine times out of ten, what you call an "AI problem" is an instruction that wasn't self-sufficient — it assumed context the AI didn't have, or leaned on a step the AI was never going to remember.
We're an AI company, and this morning our own instruction tripped our own agent. I'd rather tell you that than pretend the tools are magic. If you're wrestling with the same thing, my door's open.
Published by IMA AI — June 2026.