research-prompt-crafter

name

research-prompt-crafter

description

Craft high-quality prompts for deep research models (GPT-5.5, Gemini 3 Pro, Claude Opus 4.7, DeepSeek). Use when user wants to create a prompt for a research task that will run on a "thinking" model taking 5-20+ minutes. Triggers: "help me write a prompt", "craft a research prompt", "I need a prompt for deep research", "prompt for GPT-5.5", "prompt for Gemini", "write a prompt that will take a while to run", or any request to create prompts for long-running AI research tasks.

Research Prompt Crafter

Create prompts optimized for deep research models—the kind where you send a prompt and wait 5-20+ minutes for a comprehensive response. Getting the prompt right matters when compute time is significant.

Environment Detection

Claude Code: Use Ask User Question tool for iterative refinement. Save output as .md file. Claude Desktop / claude.ai: Use conversation for refinement. Save output as markdown artifact.

Workflow

- [ ] Step 1: Confirm target model
- [ ] Step 2: Gather brain dump
- [ ] Step 3: Clarify ambiguities (iterative)
- [ ] Step 4: Propose structure & confirm
- [ ] Step 5: Generate final prompt
- [ ] Step 6: Save output

Step 1: Confirm Target Model

Ask the user which model the prompt will run on:

• GPT-5.5 → See gpt-5-5-patterns.md
• Gemini 3 Pro → See gemini-3-patterns.md
• Claude Opus 4.7 → See claude-4-7-patterns.md
• Generic (multi-model) → Use common patterns that work across models
• DeepSeek → Use generic patterns (specific guide pending)

Load the relevant reference file before proceeding.

Step 2: Gather Brain Dump

Ask the user to provide their rough/unstructured input. Prompt for:

• Topic/question: What research question or topic?
• Scope: How broad or narrow? Any boundaries?
• Depth: Surface overview or exhaustive deep-dive?
• Output format: Report structure, length expectations, sections needed?
• Audience: Who will read this? Technical level?
• Sources/constraints: Preferred sources? Things to avoid? Time bounds?
• Success criteria: What makes a good answer?

Accept messy input—that's the point. User shouldn't need to structure this themselves.

Step 3: Clarify Ambiguities (Iterative)

Review the brain dump for:

1. Contradictions: Conflicting requirements (e.g., "brief but comprehensive")
2. Vague terms: Undefined scope words ("recent", "major", "key")
3. Missing constraints: Unstated assumptions about length, format, depth
4. Implicit requirements: Things user probably wants but didn't say

Use Ask User Question (Claude Code) or direct questions to resolve each issue. Ask 1-3 focused questions per round. Continue until requirements are clear.

Example clarifications:

• "You mentioned 'recent developments'—what time range? Last 6 months? Year? 5 years?"
• "You want both breadth and depth—if forced to choose, which matters more?"
• "Should the output include citations/sources, or just synthesized findings?"

Step 4: Propose Structure & Confirm

Based on gathered requirements, propose:

1. Prompt structure: What sections/blocks the prompt will have
2. Key instructions: Main behavioral directives for the model
3. Output specification: What the model should produce
4. Constraints: What to avoid, length limits, format rules

Present this as a brief outline and ask for confirmation or adjustments. This catches misunderstandings before writing the full prompt.

Step 5: Generate Final Prompt

Apply model-specific patterns from the reference files:

Common elements across models:

• Clear role/persona (if needed)
• Explicit task description
• Constraints and boundaries
• Output format specification
• Quality criteria

Use XML structure for complex prompts—all three major models handle XML well:

<role>...</role>
<context>...</context>
<task>...</task>
<constraints>...</constraints>
<output_format>...</output_format>

Include current date at the top of the prompt for time-sensitive research, except for GPT-5.5, which is already aware of the current UTC date — adding it just adds noise (only inject a date there for non-UTC, business, or "as-of" anchoring).

Step 6: Save Output

Filename format: {topic-slug}-{model}-prompt.md
Examples: ai-regulation-gpt55-prompt.md, market-analysis-gemini3-prompt.md

Claude Code locations (check in order):

1. design/ folder if it exists
2. Repository root otherwise

Claude Desktop: Create as markdown artifact.

Quality Checklist

Before finalizing, verify:

• Task is unambiguous—model won't need to guess intent
• Scope is explicit—clear boundaries on what to include/exclude
• Output format is specified—structure, length, sections
• Constraints are stated—what NOT to do
• Success criteria defined—what makes a good response
• Model-specific patterns applied from reference file

Anti-Patterns to Avoid

• Vague scope: "Tell me about X" without boundaries
• Contradictory instructions: "Be brief but cover everything"
• Missing format spec: No guidance on structure/length
• Assumed context: Referencing things the model won't know
• Over-instruction: So many rules the model gets confused