kai/knowledge-base
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
e61baf7e4e3032fd805d4497117f53d6ca042ae9
knowledge-base/6 - Zettelkasten/20260320100100 上下文腐烂与全新窗口隔离.md
Yaojia Wang e61baf7e4e vault: add GSD methodology deep-dive with examples and zettelkasten insights 
Comprehensive GSD analysis: 15 sections covering core philosophy (fresh
context per agent), 5 methodologies (dream extraction, goal-backward
verification, nyquist validation, wave execution, checkpoints), full
command reference (37+), agent system (16 agents with model routing),
config system, git integration, state management, session continuity,
community best practices, pitfalls, framework comparison (GSD vs ECC vs
BMAD vs SpecKit), and 4 detailed practical examples (new project, brownfield,
debugging, quick tasks).

Three zettelkasten notes: context rot vs window isolation tradeoffs,
goal-backward vs forward verification, plans-as-prompts design pattern.
2026-03-20 00:19:23 +01:00

1.0 KiB
Raw Blame History

created, type, tags, source
created type tags source
2026-03-20 10:01 zettel
claude-code
context-window
ai-quality
gsd
https://github.com/gsd-build/get-shit-done

上下文腐烂与全新窗口隔离

AI 编码 Agent 的输出质量随上下文填充率非线性下降:

  • 0-30%: 峰值质量
  • 50%+: 开始赶工
  • 70%+: 幻觉风险显著增加

GSD 的解决方案是激进隔离: 每个任务执行器获得全新的 200K token 上下文窗口，编排器保持在 30-40% 使用率。这与 ECC 的"策略性压缩"（在逻辑节点手动 /compact）形成对比。

两种方法的权衡:

  • 全新窗口: 质量最佳，但 token 成本高（每个 Agent 独立 API 调用）
  • 策略性压缩: 成本低，但依赖人类判断何时压缩，且压缩可能丢失关键上下文

理想方案可能是两者结合: 对复杂多阶段项目用 GSD 的窗口隔离，对日常编码用 ECC 的策略压缩。

Related