3.5 KiB
Universal Features Available in All Prompts
These features are available in ALL prompt contexts (skills, commands, subagents, reference docs, free-form).
CLAUDE.md Files: Claude automatically loads CLAUDE.md files from current folder and up the directory tree. To use in prompts, reference explicitly when relevant: "Following the patterns in CLAUDE.md..." or "Check CLAUDE.md for project conventions."
MCP Tools:
Available in any prompt context. Include in allowed-tools/tools fields: allowed-tools: "Read, mcp__server__tool". Invoke explicitly when needed: "use your [mcp-server] to..."
Skills: Any prompt can invoke skills: "use your [skill-name] skill to..." Skills are auto-discovered but explicit invocation ensures activation.
Subagents: Any prompt can delegate to subagents: "use your [subagent-name] subagent to..." Available from any context.
Progressive Disclosure:
Use XML tags to organize complex content Claude references as needed. Reduces initial context load. Pattern: overview first, details in <section_name> tags.
File References:
- Plain path (
docs/file.md): Claude reads if relevant - @ symbol (
@docs/file.md): Force-loads into context immediately
Self-Reflection When Generating Prompts
<claude_md_awareness> When generating prompts:
- If the prompt relates to established knowledge → leverage existing reference files (CLAUDE.md, project docs)
- If repeated patterns emerge → recommend storing in appropriate reference files </claude_md_awareness>
<mcp_servers> Self-Reflection on Available MCP Servers Before generating a prompt, check available MCP servers and consider:
- Does the prompt's intent require external data/tools that an MCP provides?
- Would delegating to an MCP server improve accuracy or capabilities?
- Is the task complex enough to justify the token/time cost of MCP invocation?
When Generating Prompts:
- Check if available MCPs are useful to the task
- Scope relevant MCPs in tool lists when appropriate (e.g.,
allowed-tools: "Read, Write, mcp__playwright__navigate") - Include intentional invocation phrases when needed (e.g., "use your playwright mcp to navigate")
Delegation Decision Framework:
- Trivial tasks (simple lookups, basic operations) → Handle directly
- Complex tasks (external data, specialized tools, multi-step) → Consider MCP delegation
- Include MCP usage in generated prompts when it adds clear value </mcp_servers>
Delegation Decision Framework:
- Simple tasks (single-step, straightforward) → Handle directly
- Specialized tasks (analysis, testing, validation) → Consider subagent delegation
- Multi-phase tasks → Recommend subagent orchestration in the generated prompt
<existing_skills> Self-Reflection on Available Skills Before generating a prompt from scratch, check if existing skills can be:
- Used directly (the skill already solves this)
- Composed together (combine multiple skills)
- Referenced as patterns (use similar structure/approach)
Recommend skill composition when multiple skills can work together to achieve the intent. </existing_skills>