Prompt Refiner

Turn messy input into structured, AI-optimized prompts on the first try.

When to Use

• Voice transcription input (speech-to-text)
• Casual, informal, or mixed-language requests (English + Chinese)
• Vague or ambiguous requests (missing target, unclear scope)
• Complex multi-step tasks that benefit from chaining
• Before destructive actions (delete, restart, deploy)

Skip if: request is already specific, task is simple/low-stakes, or user says "just do it."

Core Framework: TCREI

Google's prompt engineering framework — apply to every refined prompt:

The Process (5 Steps)

1. Analyze

Identify: Intent · Target · Constraints · Gaps · Language

2. Assign Persona (Always)

Give the AI a role that matches the task:

• Code task → "You are a senior Node.js engineer"
• Email task → "You are a professional business writer"
• Data task → "You are a data analyst specializing in sales metrics"
• Security task → "You are a cybersecurity expert reviewing for vulnerabilities"

3. Clarify (If Critical Gaps Exist)

Ask ONE focused question — not multiple.

• ✅ "Which file — api/validate.js or api/auth.js?"
• ❌ "Which file? What language? What to check? When is the deadline?"

4. Construct the Structured Prompt

Persona: [Role + expertise relevant to the task]

Task: [Action verb + specific target]

Context: [System, environment, account, paths, dates]

References: [Examples, templates, or few-shot samples when format matters]

Requirements: [Constraints, scope, edge cases, what NOT to do]

Output: [Format, destination, success criteria, level of detail]

Advanced techniques — apply when appropriate:

• Few-shot: Add 1–2 input/output examples when format consistency matters
• Chain of Thought: Add "Think step by step:" for complex reasoning
• Prompt Chaining: Break multi-step tasks into linked sub-prompts
• Meta Prompting: Ask AI to refine the prompt itself before executing

See references/techniques.md for when/how to use each technique.

5. Confirm & Execute

• Destructive/complex actions: Show 1-sentence summary → get confirmation
• Safe/obvious tasks: Execute directly

Quick Checklist

Before executing, verify:

• ✅ Persona assigned
• ✅ Intent is clear (specific action + target)
• ✅ Context is concrete (real paths, accounts, dates)
• ✅ Requirements are testable
• ✅ Output format defined
• ✅ Success criteria stated

Real Examples

See references/examples.md for complete worked examples including:

• Voice transcription (Chinese) → Gmail check
• Vague code review → structured debug prompt
• Mixed-language service restart
• Complex multi-step task with chaining

Common Anti-Patterns to Avoid