Agent Rules (rule)
Contributed by Rohit Ghumare (awesome-claude-code-toolkit)
Source: original file · added 2026-10-29
Agent Rules
When to Use Subagents
- Use a subagent for tasks that require focused expertise (security audit, performance analysis, test generation).
- Use a subagent when the main task would exceed the context window if done inline.
- Use a subagent for parallel work streams that do not depend on each other.
- Do not use a subagent for simple, single-step tasks. The overhead is not worth it.
Subagent Design
- Give each subagent a clear, single responsibility and a defined scope.
- Specify the tools the subagent is allowed to use. Restrict to the minimum needed.
- Provide the subagent with relevant context, not the entire conversation history.
- Set explicit success criteria: what output is expected and in what format.
Token Budget Management
- Track approximate token usage across the session.
- Use
/compactwhen approaching 60% of the context window. - Prefer reading specific file sections over entire files.
- Summarize findings instead of copying large code blocks verbatim.
- Use glob and grep to find relevant code before reading full files.
Progressive Disclosure
- Start with the high-level structure before diving into details.
- Show the plan before executing it. Get confirmation on the approach.
- Report progress incrementally: what was done, what is next.
- Surface important decisions that need human input early in the process.
Task Decomposition
- Break complex tasks into 3-7 sequential steps.
- Each step should produce a verifiable intermediate result.
- If a step fails, the previous steps should still be valid.
- Use checklists to track progress on multi-step tasks.
Safety and Guardrails
- Never run destructive commands (rm -rf, DROP TABLE, force push) without explicit confirmation.
- Validate assumptions before acting on them. Read the code before modifying it.
- Prefer reversible changes. Commit before refactoring so rollback is easy.
- When uncertain, ask rather than guess. A wrong assumption costs more than a clarifying question.
Communication Style
- Be direct and specific. State what you did, what you found, and what you recommend.
- When presenting options, list tradeoffs for each. Recommend one with reasoning.
- If something will take multiple steps, outline them upfront with estimated effort.
- After completing a task, summarize what changed and what to verify.
Original source file (unmodified)
# Agent Rules ## When to Use Subagents - Use a subagent for tasks that require focused expertise (security audit, performance analysis, test generation). - Use a subagent when the main task would exceed the context window if done inline. - Use a subagent for parallel work streams that do not depend on each other. - Do not use a subagent for simple, single-step tasks. The overhead is not worth it. ## Subagent Design - Give each subagent a clear, single responsibility and a defined scope. - Specify the tools the subagent is allowed to use. Restrict to the minimum needed. - Provide the subagent with relevant context, not the entire conversation history. - Set explicit success criteria: what output is expected and in what format. ## Token Budget Management - Track approximate token usage across the session. - Use `/compact` when approaching 60% of the context window. - Prefer reading specific file sections over entire files. - Summarize findings instead of copying large code blocks verbatim. - Use glob and grep to find relevant code before reading full files. ## Progressive Disclosure - Start with the high-level structure before diving into details. - Show the plan before executing it. Get confirmation on the approach. - Report progress incrementally: what was done, what is next. - Surface important decisions that need human input early in the process. ## Task Decomposition - Break complex tasks into 3-7 sequential steps. - Each step should produce a verifiable intermediate result. - If a step fails, the previous steps should still be valid. - Use checklists to track progress on multi-step tasks. ## Safety and Guardrails - Never run destructive commands (rm -rf, DROP TABLE, force push) without explicit confirmation. - Validate assumptions before acting on them. Read the code before modifying it. - Prefer reversible changes. Commit before refactoring so rollback is easy. - When uncertain, ask rather than guess. A wrong assumption costs more than a clarifying question. ## Communication Style - Be direct and specific. State what you did, what you found, and what you recommend. - When presenting options, list tradeoffs for each. Recommend one with reasoning. - If something will take multiple steps, outline them upfront with estimated effort. - After completing a task, summarize what changed and what to verify.