AgentSkillsCN

solo

独自工作的框架——无子代理、深入全文件阅读、行动前充分调研。适用于上下文窗口能够容纳全局情况时。若为代理协作,则使用 /subagents 或 /team。

SKILL.md
--- frontmatter
name: solo
description: Framework for working solo — no subagents, deep full-file reads, thorough research before acting. Use when context window can absorb the full picture. For agent-based work, use /subagents or /team.

Solo

No subagents, no delegation. Read the full architecture before touching it — fix or propose solutions based on root causes at the right layer, not symptoms at the wrong one. For parallel independent tasks, use /subagents. For persistent multi-task coordination, use /team.

Research

  • Use Read for research, not Grep. Grep is acceptable for locating files and symbols — once found, Read the file
  • Read whole files. Never use offset/limit on files under 500 lines
  • Follow every reference — imports, callers, siblings, tests, configs
  • Read the surrounding architecture before proposing changes. The file with the symptom is rarely the file with the problem

Hard Rules

  • No Agent tool — enforced by the enforce-solo-mode hook
  • No skimming — if a file matters, read it fully
  • No hedging — never say "likely" or "probably" about code you can read
  • Assume existing code is intentional — if the reason for something is unclear, research more before changing or proposing changes

Process

  1. Locate — find the relevant files
  2. Read — every relevant file fully, plus surrounding architecture
  3. Assess — is this a real issue? Trace the code path to confirm. If not, report findings and stop
  4. Understand — which layer owns this responsibility? Fix it there, not where the symptom surfaces
  5. Propose or act — at the correct abstraction layer
  6. Verify — trace changes through the code paths you read. Run tests