How I Deleted 95% of My Agent Skills and Got Better Results — Nick Nisi, WorkOS
TL;DR · AI Summary
Nick Nisi found that by trimming AI agent skills from 95% down to just 5 core roles (implementer, verifier, reviewer, closer, retro), he achieved higher-quality outputs; the key was replacing Claude-native skills with a TypeScript state machine to solve context loss.
Key Takeaways
- Reduced agent skills from 95% to 5%, retaining only five roles (implementer/veri
- After refactoring with a TypeScript state machine, context loss and task skippin
- The new system auto-extracts context from GitHub issues, PRs, Slack threads, or
Outline
Jump quickly between sections.
As a DX engineer maintaining 20+ repos across 8 languages, the author spent ~10 minutes per task manually re-establishing context for single-agent workflows.
The first ‘Case’ tool used Claude skills to ingest GitHub/Slack/Linear inputs and execute tasks, but failed under complexity due to context drop and instruction ignoring.
A TypeScript state machine now orchestrates five dedicated agents (implementer, verifier, reviewer, closer, retro), ensuring traceable, evidence-backed execution.
After deleting 95% of redundant skills, the system became more reliable, delivered higher-quality PRs, and reduced human intervention—proving ‘less is more’ in agent design.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- 精简AI代理技能提升效能
- 问题根源
- 多仓库上下文切换成本高
- 每次需10分钟人工重建背景
- 单代理易遗忘/跳步
- 解决方案
- Case工具：输入→自动上下文提取
- TypeScript状态机编排
- 五角色分工：Implementer/Verifier/Reviewer/Closer/Retro
- 效果验证
- 技能删减95%，质量反升
- PR必须含可验证证据
- 人工干预减少，交付更快更稳
Highlights
Key sentences worth saving and sharing.
I deleted 95% of my agent skills, keeping only five core roles—implementer, verifier, reviewer, closer, and retro—and got better results.
The initial Claude-based Case worked well until complexity increased; then it started forgetting instructions or skipping steps—for example, when asked ‘Why didn’t you do that?’, it replied ‘Oh yeah,
The new system automatically parses GitHub issues, PRs, Slack threads, or Linear tickets—no manual context setup—and requires agents to produce PRs with verifiable evidence before stopping.