There is a version of the “more is better” instinct that feels entirely reasonable when you are building with AI. You have a capable model, you have hard-won lessons about what goes wrong, so you write them all down. You refine. You add the edge cases. By the time you are done, your instructions are thorough, and you feel good about that. Then the model starts missing delivery.
A structured experiment with Fable 5 — Anthropic’s latest reasoning model — reframes what instructions are actually for.
The experiment
The experiment involved auditing an AI harness — the full layer of accumulated instructions, rules, skills, and files that surrounds a model session. The audit produced a dense method: 5,197 words covering how to inspect what is loaded, how to map what shapes a given run, how to assign each control to one of six dispositions. It was good work. A compact version was also available: 742 words, four required elements, same job.
Both were run against the same content task at maximum reasoning effort and the outputs were scored blind.
The full method won on quality. Fable 5 working through 5,000 words of structured guidance scored 19.67 out of 20. The compact brief scored 17.5. When the model had more to think with, it thought better. That part matched my intuition.
But there was also a delivery contract: the output had to arrive as valid JSON, under 1,400 words. Concrete, binary requirements.
The compact brief passed that contract three times out of three. The full method passed it once.
What actually broke
The failure mode was not that the model ran out of capacity or got confused. The full method and the delivery requirements were competing for the same attention, and that competition had not been made explicit.
The method was rich and reasoning-heavy — it told the model how to think its way through a complex audit. The delivery requirements were buried inside that prose as reminders. They were not structured differently. They did not announce themselves as hard constraints. From the model’s point of view, they were just more text in a long document.
When you give a model a sophisticated method to reason through, it applies effort to the reasoning. That is what you asked it to do. If your enforcement requirements are written in the same register as your guidance, the model has no obvious signal that one class of instruction is advisory and the other is binary. It tries to serve both, and the binary ones are the ones that show up as failures when they are missed.
This is not a model problem. It is a harness design problem.
The right combination matters more than the length
There is a tempting conclusion here — “long prompts are bad” — and it is worth resisting, because the evidence does not support it. The 5,000-word method contained real value. The model’s output quality showed that. OpenAI’s internal research on coding agents found that leaner prompts improved evaluation scores roughly 10 to 15 percent and cut costs significantly, with the caveat to remove one instruction group at a time and keep examples that encode product requirements. That is not “shorter is better”; it is “precision matters.”
The lesson is narrower and more actionable: do not mix rich reasoning guidance with prose enforcement. When a requirement is binary — valid JSON or not, under a word count or not — it should not live in a paragraph. It should be structurally distinct, or it should be enforced outside the prompt entirely, through a validator that checks the output after the fact.
Put another way: give judgment to prose, give enforcement to code.
What this changes about how to build
Instructions have two distinct jobs. One job is to help the model understand the problem, the context, and the method — and for that, richness and nuance genuinely help, especially with a model built for extended reasoning. The other job is to enforce contract requirements — and for that, prose is the wrong tool. Prose can be interpreted. A downstream check cannot.
As models get more capable at reasoning, the temptation to load them with more sophisticated guidance will grow. That is probably fine. But the delivery layer needs to keep up, and it does not do that by getting longer. It does that by getting more structural.
The harness that works well today is one that knows which of its parts are teaching the model to think, and which parts are ensuring the work actually lands.
Text summarized and optimized using Anthropic’s models and reviewed by a human.