41 lines
1.4 KiB
Markdown
41 lines
1.4 KiB
Markdown
# Decision Framework
|
|
|
|
## Start Here: Natural Language
|
|
|
|
Every prompt should begin with clear, direct natural language instructions. This is your baseline.
|
|
|
|
## Then Ask: Would a prompting technique help?
|
|
|
|
- Complex reasoning needed? Consider CoT
|
|
- Need specific format/style? Consider Few-Shot
|
|
- Multiple types of information? Consider XML separation
|
|
- Already clear and effective? Stop, you're done
|
|
|
|
## Then Ask: Would Claude Code features help in the prompt content?
|
|
|
|
- Should this prompt reference CLAUDE.md knowledge?
|
|
- Should this prompt invoke existing skills?
|
|
- Should this prompt delegate to subagents?
|
|
- Should this prompt invoke MCP tools?
|
|
- Complex structure? Use XML tags for progressive disclosure
|
|
|
|
## Principle: Less is More
|
|
|
|
Don't add techniques just because you can. Add them when they solve a specific problem. Most prompts work best with just clear natural language.
|
|
|
|
## Skill Invocation Patterns
|
|
|
|
For skills, include trigger phrases to ensure intentional activation:
|
|
|
|
- "use your [skill-name] to..."
|
|
- "use the [skill-name] for..."
|
|
- Include trigger words from skill description
|
|
|
|
For subagents, use clear delegation language:
|
|
|
|
- "use your [subagent-name] subagent to..."
|
|
- "use the [subagent-name] subagent for..."
|
|
- Match language to subagent's description triggers
|
|
|
|
Claude auto-discovers skills and subagents but may skip them without explicit triggers in the prompt.
|