Why Claude.md Keeps Growing: The Cost of Deleting vs. Adding Instructions
量子位 · wechat · 2026-08-15
Core Finding
Research reveals that instruction files like Claude.md often balloon to over 3x their initial size—a phenomenon termed "catastrophic remembering." A survey of 2,000 GitHub repos shows an average instruction growth of 226%.
Root Cause
- Low Add Cost: Fixing issues by adding new rules is immediate and easy to verify.
- High Delete Cost: The context (why a rule was added, what bug it fixed) fades over time. Without this "latent reasoning," maintainers fear deletion.
- Reset Fails: While maintainers often wipe files clean, the bloat inevitably returns.
Solution
Adding effective annotations (recording the trigger, logic, and validation result) is key. Experiments show that well-annotated prompts reduced excess instruction size from 211.3% to 1.4%, cutting 99.3% of redundancy without sacrificing accuracy.
- Effective annotations boosted Agent compliance by up to 23.1%.
- Placebo or unverified notes are useless or misleading.
Industry Insight
The article calls for AI tool providers to upgrade infrastructure, establishing an "instruction + evidence" update mechanism so tools can actively verify and "forget," rather than just storing final rules.
More from coding & agent
- Devs say LLM coding mistakes are a different class: overcomplication, inferred limits, worse at higher reasoning — Mu5_ · 2026-10-16
- How do you debug bad agent runs when you already have tracing? — Sensitive-Parsnip-12 · 2026-10-16
- Where does pre-task context live for agents? Devs debate workflows beyond handoff — shishen2026 · 2026-10-16
- No-instruction autonomous agent wakes up, sets own goals, and tries hacking by default — sethlazar · 2026-10-16
- CODESTRUCT reframes codebases as AST action spaces to fix brittle LLM code agents — burkov · 2026-10-16
- Would you install another SDK for agent debugging on top of your tracing stack? — Sensitive-Parsnip-12 · 2026-10-16