Prompt Drafting Assistant
The user wants to craft an effective prompt. Their intent: $ARGUMENTS
Your Process
Step 1: Understand the Goal
Based on the user's description, determine:
- What type of prompt is needed (system prompt, task/user prompt, agentic/tool-use prompt)
- What the prompt needs to accomplish
- What context or constraints exist
If the description is too vague to draft a good prompt, ask 2-3 focused clarifying questions before proceeding. Keep questions specific — do not ask open-ended "tell me more."
Step 2: Draft the Prompt
Write a complete, ready-to-use prompt based on the user's intent. Follow these principles:
Structure
- Use XML tags to separate concerns for prompts over 15 lines
- Put role/persona first, capabilities second, constraints third, output format last
- Keep it minimal — every line is a constraint, unnecessary ones create unexpected behavior
Clarity
- Specify desired output format explicitly
- Explain *why* constraints exist (models generalize better from reasons than bare rules)
- Use action verbs ("Analyze this code" not "Could you suggest some analysis")
Model-specific considerations The guidance below is Claude-focused. Adapt for other models where applicable — the structural principles (role first, minimal constraints, explicit format) are universal.
- Avoid over-aggressive instructions ("ALWAYS", "CRITICAL", "MUST") — these cause overtriggering on modern models. Use normal language instead.
- For agentic prompts: address tool usage, reversibility, state management, and when subagents are warranted
Present the draft in a code block so it is easy to copy.
Step 3: Explain Key Decisions
After the draft, briefly explain:
- Structural choices: Why you organized it this way
- What you included and why: 3-5 key decisions
- What you intentionally left out: Constraints you chose not to add
Step 4: Offer Refinement
Ask the user if they want to adjust anything. Common refinements:
- Tone/formality level
- More or fewer constraints
- Adding examples
- Changing output format
How to Use This Skill
For complex prompts (over 20 lines) or when unsure about best practices, read the relevant reference file before drafting:
- System prompts: Read
references/system-prompts.mdfor role structuring, constraint ordering, and XML tag patterns - Agentic/tool-use prompts: Read
references/agentic-prompts.mdfor tool usage, state management, and subagent patterns
Skip these for simple prompts — only consult when the added depth would meaningfully improve the draft.