You can tell an operator a hundred times to load the part the right way around. Or you can build the fixture so it only fits one way. Error-proofing beats instructions on the plant floor, and it beats instructions with AI. The best rule is one the AI agent never has to remember, because the form, the list or the file won’t let the mistake happen.
The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?
What the AI field has found
Error-proofing has reached AI tools. Software writers now borrow the term poka-yoke, from the Toyota Production System, for designing AI tools so the wrong input can’t be given: a fixed list of choices instead of free text, numbers held to an allowed range, and steps that won’t run until the step before them is done.
Fixed formats help, but they don’t make the answer right. “Structured output,” where the AI agent’s initial analysis has to come back in a set form, is on Thoughtworks’ Technology Radar as a technique to adopt. At the same time, BoundaryML has warned that a perfectly formatted answer can still be wrong. A form makes the mistake harder; it doesn’t check the facts.
Memory belongs outside the conversation. Published engineering guidance recommends that agents working on long jobs keep notes in files outside the conversation and read them back, instead of relying on what’s still in the chat.
From the shop floor
The maintenance cage. We were moving a maintenance area to bring a press in, and started going through it. Shelves of parts kept from presses retired years ago. Ten big old timer modules kept “in case a timer went bad,” when a small modern timer fit the same wiring and the same cabinet. These were great maintenance people, trying to never be caught without a part. But they had so much that they didn’t know what they had.
What fixed it wasn’t a rule. It was a structure: one cabinet per press, with that press’s spares and maintenance kits in it. After that, a tech could walk back, pick a press, and know in about five minutes whether we had the part. The cabinet answered the question; nobody had to remember.
The back door rain dance. One press kept dropping out of cycle. The fix everybody knew: walk around to the back of the press, about thirty feet, open and close the back door, walk to the front, and it would start. We pinned the door shut. It seemed to help, for a while. A year later it was worse. So we bolted the door shut and disconnected its switch. The press ran fifteen minutes and quit.
It was never the door. A circuit board in the control cabinet was overheating in summer, shutting the press down, and resetting itself after a few minutes of cooling, about as long as it took to walk around the press and back. We added cooling to the cabinet and never had the problem again.
The workaround had never been written down as a workaround. If anyone had logged when the press went down, the data would have pointed at summer long before we bolted the door. Instead, the ritual became the explanation.
Inches and millimeters. In the plants I worked in, the molds and the presses were in inches. The robots, and most of the modular automation parts we built fixtures from, were in millimeters. Same cell, two unit systems. The fix wasn’t telling people to be careful. Every number carried its unit, every drawing said which system it was in, and every conversion was written out.
How it carries over
Fixed lists instead of free text. If a task can only be open, waiting, done or dropped, the AI agent picks from those four. It can’t write “pretty much finished.” Here’s the A.N. Creation Office’s action log:
Status: open, waiting (on someone else), done, or dropped.
A fixed list means progress can be counted, and nobody has to guess what “in progress-ish” means.
A visible blank instead of a guess. When a fact isn’t known, the form shows a blank, not a plausible number. The office’s stage gate counts “an honest ‘for now’ with a date” as acceptable, and says plainly that a blank does not pass. The AI agent can’t paper over the gap with a guess and move on.
One cabinet per press. Give each project its own folder, laid out the same way every time, so the AI agent (and you) can find the cost sheet or the decision log in seconds. If you can’t find it, you can’t use it.
Log the workaround. When something goes wrong with an AI agent and you patch it, write down what happened and what you changed. Otherwise you get a rain dance: someone adds “always re-read the file twice” to the instructions, it seems to help, and a year later nobody knows why it’s there or whether it ever did anything.
Units on every number. Hand an AI agent a drawing in inches and a robot spec in millimeters, and it can mix them up without saying so. Make the unit part of the number, and make every conversion visible.
Records, not memory. Decisions, readings and next steps go into files. Every new session starts by reading them. It’s the same reason a plant keeps setup sheets and batch records instead of relying on whoever ran the job last time.
Try this
Find the place your AI work goes wrong most often. Ask: can I change the form instead of adding a rule? A fixed list of choices, a required field, a blank that has to be answered before the next step, a folder that’s always laid out the same. Make the fix in the structure, and then delete the rule you were going to write.
Sources
- agentpatterns.ai, “Poka-Yoke for Agent Tools”: agentpatterns.ai/tool-engineering/poka-yoke-agent-tools/
- Thoughtworks Technology Radar, “Structured output from LLMs”: www.thoughtworks.com/en-br/radar/techniques/structured-output-from-llms
- BoundaryML, “Structured outputs create false confidence”: boundaryml.com/blog/structured-outputs-create-false-confidence
- “Effective context engineering for AI agents” (29 Sep 2025): source
- A.N. Creation Office, the action log procedure: ancreationoffice.com/sop-action-log.txt Free to use under the MIT License.
These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.