Your AI Assistant's Real Weakness Isn't Smarts — It's Its To-Do List
Messy AI instructions cause wrong answers and bigger bills — and that's fixable.
Most AI research tests the brain of the model — how smart it is. This paper looks somewhere else entirely: the folder of written instructions that tells an AI agent (software that takes actions for you) how to do specific jobs. The author calls these files "skills," and his tool, SkillSeam, deliberately breaks one thing at a time to see what goes wrong. The core finding is simple: a stack of good instructions is not automatically a good system. Skills fail at the seams where they meet each other.
The numbers are concrete. When instructions were made flatter instead of layered, the AI had to read 60% more text — which means slower answers and higher bills. A single broken link between instructions raised reading volume by 64% and knocked accuracy down 3.1 points. Renaming one instruction with a near-identical name sent wrong-task routing from 0 out of 32 cases to 15 out of 32. Vague trigger words were the worst offender: routing conflicts jumped from 3 in 32 to 30 in 32, and the AI read 3.7 times more text. The single biggest accuracy hit — a 12.5-point drop — came from mixing instructions that were wildly different sizes.
Why should you care? You probably use an AI assistant that reads instructions like these, whether it's booking appointments, sorting email, or answering customers. When those instructions blur together, the AI picks the wrong one, does the wrong task, or stalls — and because you pay per chunk of text processed, messiness literally costs money. It also explains a common frustration: "my AI got dumber today." Often the model didn't change; the instructions around it did.
One honest caveat: this is a single-author study on one sealed test system, so the exact percentages won't apply everywhere. The author released his test files and a one-page checklist so others can repeat the work. The practical takeaway is unglamorous but real — keep your AI's instructions clean, uniquely named, and roughly the same size, and it will behave better.
- AI assistants fail most often where instructions overlap or collide, not because one instruction is bad.
- Vague trigger wording pushed wrong-task mix-ups from 3 in 32 to 30 in 32, and made the AI read 3.7 times more text.
- The biggest accuracy drop, 12.5 points, came from mixing instructions of very different sizes.
Why It Matters
Cleaner AI instructions mean fewer wrong actions, faster replies, and lower per-use costs for you.