本文由 AI 分析生成
建立時間： 2026-05-23 來源： https://x.com/servasyy_ai/status/2056611371022954651
Summary
A Chinese-language post citing Karpathy’s claim that 90% of Claude errors come from missing context rather than model capability. It presents 12 engineering rules distilled by senior engineers that reduce error rate from 41% (no CLAUDE.md) to 3% (12 rules). Rules cover thinking before coding, surgical edits, goal-driven execution, token budgeting, exposing conflicts, and loud failure signaling.
這篇文章引用 Karpathy 的說法：Claude 90% 的錯誤來自上下文缺失而非模型能力不足。呈現 12 條資深工程師整理的規則，可將錯誤率從 41%（無 CLAUDE.md）降至 3%。涵蓋先思考再寫碼、外科手術式修改、目標驅動、Token 預算、顯式暴露衝突、大聲失敗等原則。
Key Points
- No CLAUDE.md: 41% error rate; 4 basic rules: 11%; 12 rules: 3%
- Rule highlights: think before coding (state assumptions), minimal code (no speculative abstraction), surgical edits only, goal-driven with success criteria, use model for judgment tasks only, token budgets (4k/task, 30k/session), expose conflicts (never average them), read before write, test intent not behavior, checkpoint every step, match codebase conventions, fail loudly
- CLAUDE.md as cross-session organizational memory is the highest-leverage investment
Insights
The 41% → 3% error rate curve from context engineering alone is striking and aligns with the broader insight that harness/context quality dominates raw model capability for practical tasks. Rule 5 (use model for judgment tasks only, not routing/retry/deterministic transforms) is particularly actionable — it defines a clear boundary for when to involve LLMs vs. conventional code. Rule 12 (loud failure over silent skips) addresses a failure mode that is uniquely dangerous in LLM-generated code.
Connections
Raw Excerpt
没有 CLAUDE.md 时，错误率是 41%。有 4 条基础规则时，错误率降到 11%。使用下面这套 12 条规则时，错误率降到 3%。真正会复利增长的，不是下一个框架，而是：把 CLAUDE.md 当作跨会话的组织记忆。