The Experiment That Should Change How You Set Up AI
Nate Jones ran a clean test: take Fable 5, add roughly 5,000 words of carefully written instructions, and measure what happens. Analysis scores improved. Delivery failed two out of three runs.
That's not a paradox — it's a system design problem. And it's almost certainly happening inside your AI workflows right now.
The culprit Jones identified: harness bloat. In one common workflow, the model was processing 18,384 words of accumulated rules before it ever reached the actual task instructions. Every patch, every fix, every "just add a rule for that" decision compounds invisibly until the model is spending most of its context window on setup overhead instead of doing the work.
Why Operators Miss This
Most teams never audit their AI harness because it doesn't look broken. The model still responds. It still sounds coherent. The failure shows up as inconsistency — outputs that are sometimes great, sometimes wrong, with no obvious pattern. Teams blame the model or the prompt when the real issue is architectural.
This is especially common in non-tech companies adopting AI without dedicated ML infrastructure. You start with a simple system prompt. Someone adds edge case handling. A consultant recommends a few more guardrails. Six months later you have a bloated instruction stack that nobody owns and nobody has read end-to-end.
The insight from Jones's experiment is precise: more context can improve reasoning quality while simultaneously degrading reliable execution. A model that "understands" better but fails to complete the task is not a better model for production use.
The Audit You Should Run This Week
Jones published two runnable skills — a Claude Edition and a Codex Edition — that map your existing AI setup, flag what's consuming context unnecessarily, and generate a cleanup plan before anything is changed. That last part matters: you get a receipt of the current state before you touch it.
The practical action for any operator:
- Inventory your harness. Pull together every system prompt, instruction file, and rule document feeding your highest-usage AI workflows. Count the words. Most teams are surprised.
- Separate reasoning support from execution instructions. Context that helps the model analyze is not the same as context that helps it deliver. Conflating the two creates the failure mode Jones documented.
- Run the audit tools before your next prompt revision. Optimize structure first, content second.
The lesson isn't "write shorter prompts." It's that AI setup has the same maintenance debt as any other system — and it compounds faster than most operators expect.