Your new model is running on instructions you wrote for the old one. Grab the skill and the guide: the map, the six rules, the cleanup, and the receipt.

Source: Nate’s Newsletter Date: 2026-07-15 URL: https://natesnewsletter.substack.com/p/ai-harness-audit

Summary

Nate’s Newsletter pitches an “AI harness audit” — a paid diagnostic (two runnable skills, Claude and Codex editions) for auditing the accumulated configuration around a model: custom instructions, project files, memory, skills, permissions, examples. Core claim: harnesses grow by accretion (one correction at a time) until they’re actively counterproductive — the author’s own setup hit 18,384 words before reaching platform-specific guidance, and in testing, 5,000 extra words of “good” instructions caused delivery failures 2 of 3 times versus 3-of-3 success with a compact brief. Ships with six maintenance rules and an approval-gated cleanup plan.

Implications

• Feeds the agentic engineering patterns thread directly: an empirical (if vendor-self-reported, unreplicated) data point that instruction volume has negative marginal returns and can actively degrade agent reliability as the underlying model improves — “the engine got smarter, the harness didn’t.”
• Relevant to this project’s own hygiene: landscape docs are already rewritten-in-place rather than appended for exactly this reason; this signal argues the same discipline applies to CLAUDE.md and skill/instruction files generally, and that periodic harness audits (not just landscape docs) are worth doing deliberately rather than assuming accreted rules stay net-positive.
• Read skeptically — it’s a newsletter selling a product off this claim, single-source, no methodology detail on the 2-of-3/3-of-3 test. Worth tracking as a pattern, not citing as proven.