The Problem of Configuration Bloat

Coding agents often suffer from performance degradation due to bloated configuration files (e.g., CLAUDE.md, AGENTS.md). Developers frequently add rules to these files in response to specific errors, leading to "knowledge base" files that exceed recommended lengths (often 200+ lines). Research indicates that this context bloat, along with skill and lint leakage, is widespread. Anthropic’s own experience with Claude 5 generation models showed that removing over 80% of the system prompt resulted in no measurable loss in coding performance, suggesting that much of our manual steering is redundant or outdated.

Rethinking Skills and Context

While personalized skills are popular, recent research suggests they are not a silver bullet. A study found that personalization based on individual history performed no better than generic community-sourced skills.

  • Context Files: These are most effective when they provide information the model cannot infer from the code itself, such as project-specific safety rules, expensive command warnings, or architectural boundaries. They are poor substitutes for actual code; the model is significantly more likely to answer behavioral questions by reading the source code directly rather than a prose summary.
  • Hygiene: Treat your agent configuration as a decision guide rather than a comprehensive manual. If a rule is critical, encode it into a test, hook, or permission rather than relying on prose instructions that the model may ignore or misinterpret.

Practical Audit Strategy

To maintain agent performance, adopt a regular cadence (every 2–4 weeks) for auditing your environment:

  1. Run Diagnostic Tools: Use built-in commands like /doctor (in Claude Code) to identify unused skills, plugins, and configuration cruft. Note that /doctor inside a session is for configuration, while shell-level commands may only check installation status.
  2. Review Memory: Use /memory separately from your configuration files, as auto-memory can retain stale preferences even after project files have been cleaned.
  3. Test Removal: Periodically attempt to complete tasks without local skills to see if the model and harness have improved enough to handle the work natively. This helps overcome the "crutch" mentality where developers keep outdated skills out of fear of quality drops.
  4. Prioritize Quality: When evaluating skills, focus on those with clear descriptions, triggers, and concrete examples. Use community tools to audit the quality of SKILL.md files rather than simply accumulating more of them.