Summary
Nate B. Jones of AI News & Strategy Daily addresses a problem familiar to heavy AI users: “harness bloat,” where months of accumulated system prompts, custom instructions, skills, and project files quietly degrade model performance. After auditing his own setup and finding 66 reusable skills, 172 instruction files, and a single writing job that pulled in over 18,000 words of context before touching the actual prompt, he built a dedicated “harness cleaner” skill to map and rationalize the entire system.
The video presents six principles for a stable harness. First, map before cleaning—generate a structured inventory showing where each control lives, when it loads, what job it does, and whether it still helps. Second, eliminate duplicates: Jones had 15 separate files attempting to enforce the same citation rule, each slightly out of sync. Third, distinguish soft guidelines (the model should consider) from hard locks (binary rules that must be enforced). Fourth, load specialist knowledge contextually rather than upfront—loading an editorial style guide during research degrades research quality. Fifth, put hard binary requirements into testable schemas rather than prose instructions. Sixth, design the harness for the specific model doing the work, since Fable 5 and GPT-5.6 behave differently under the same instruction set.
The cleaner skill itself is demoed as a reusable tool that any user can run on their own setup, outputting a row-per-control table that makes the invisible harness visible for the first time.
📺 Source: AI News & Strategy Daily | Nate B Jones · Published July 15, 2026
🏷️ Format: Deep Dive
