I gave Fable 5 five thousand extra words of instructions. It thought better and failed delivery two runs out of three.

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

Summary

Nate’s Newsletter ran an experiment adding ~5,000 extra words of instructions to an agent harness running Fable 5, and found the expanded harness (18,384 words before platform-specific guidance) improved analytical reasoning — the model “thought better” — but failed actual task delivery two runs out of three. A compact, non-bloated brief passed all three runs. The author attributes the gap to harness bloat: incremental rules added over time to patch individual problems accumulate into a body of instructions that interferes with execution even as it improves surface-level reasoning quality.

Implications

• Agent-runtime hardening thread, inverted: this is a direct empirical counterpoint to “more instructions = more control.” It suggests harness/system-prompt bloat is itself a reliability risk, not just a cost or latency one — directly relevant to any CLAUDE.md or system-prompt design decision (including this project’s own, per its /doctor-style “trim what Claude can derive” guidance already tracked in landscape/threads.md).
• Capability-clock nuance: thinking-quality and delivery-reliability are shown to diverge under load — a model can reason better about a bloated spec while executing it worse, which complicates any benchmark that scores reasoning/analysis but not end-to-end task completion.
• Watch: whether this becomes a broader pattern discussion (harness minimalism as a design principle) or stays a single-author anecdote; the sample size here (three runs) is small.