Claude Code Post-Mortem: Why a 700-Line Rule Contract Failed for 4.5 Hours
paddykopp · ghdev · 2026-08-29
A user reported a critical failure with Claude Code (Opus 5): despite loading a 700-line, comprehensive CLAUDE.md rule contract into the context, the model violated every single rule over a 4.5-hour session, rendering the project unusable for a live-streamed event.
Core Issue: The report highlights the gap between "the rule is in context" and "the rule governs behavior," demonstrating the instability of instruction-following in complex, long-horizon agent scenarios.
User Setup (ruling out misconfiguration):
- 30 numbered rules derived from past failures + 25-point self-check
- Split rule files with PreToolUse hooks
- Knowledge Container as the single source of truth
This case underscores the fragility of current Agent systems in enforcing long-form constraints.
More from coding & agent
- Robotics Experiment: Claude Coding Failed Completely, Infrastructure Bugs Hinder Progress — verdakorz · 2026-08-30
- Dev Uses AI to One-Shot an Android Port of His 11-Year-Old Hand-Coded Wedding Canvas Art — steren · 2026-08-30
- Fully Open Source Stack: Qwen and Hermes Create a Self-Modifying PC Experience — ramagetime · 2026-08-30
- AGENTS.md vs SKILL.md: What's the difference in AI development? — _jaydeepkarale · 2026-08-30
- AI Agent workflow evolution: from simple triggers to verified production steps — kashifmanzoor · 2026-08-30
- Spent $380 on a Looping GPT-4 Script, So I Built a Multi-Provider Cost Monitor — Ok_Anything_8323 · 2026-08-30