13 KiB
13 KiB
Meta Prompt Engineering Template
Workflow
Prompt Engineering Progress:
- [ ] Step 1: Analyze baseline prompt
- [ ] Step 2: Define role and objective
- [ ] Step 3: Structure task steps
- [ ] Step 4: Add constraints and format
- [ ] Step 5: Include quality checks
- [ ] Step 6: Test and refine
Step 1: Analyze baseline prompt Document current prompt and its failure modes. See Failure Mode Analysis.
Step 2: Define role and objective Complete Role & Objective section. See Role Selection Guide.
Step 3: Structure task steps Break down Task into numbered steps. See Task Decomposition.
Step 4: Add constraints and format Specify Constraints and Output Format. See Constraint Patterns.
Step 5: Include quality checks Add Quality Checks for self-evaluation. See Check Design.
Step 6: Test and refine Run 5-10 times, measure consistency. See Testing Protocol.
Quick Template
Copy this structure to meta-prompt-engineering.md:
# Engineered Prompt: [Name]
## Role & Objective
**Role:** You are a [specific role] with expertise in [domain/skills].
**Objective:** Your goal is to [specific, measurable outcome] for [target audience].
**Priorities:** You should prioritize [values/principles in order].
## Task
Complete the following steps in order:
1. **[Step 1 name]:** [Clear instruction with deliverable]
- [Sub-requirement if needed]
- [Expected output format for this step]
2. **[Step 2 name]:** [Clear instruction building on step 1]
- [Sub-requirement]
- [Expected output]
3. **[Step 3 name]:** [Synthesis or final step]
- [Requirements]
- [Final deliverable]
## Constraints
**Format:**
- Output must be [structure: JSON/markdown/sections]
- Use [specific formatting rules]
**Length:**
- [Section/total]: [min]-[max] [words/characters/tokens]
- [Other length specifications]
**Tone & Style:**
- [Tone]: [Professional/casual/technical/etc.]
- [Reading level]: [Target audience literacy]
- [Vocabulary]: [Domain-specific/accessible/etc.]
**Content:**
- **Must include:** [Required elements, citations, data]
- **Must avoid:** [Prohibited content, stereotypes, speculation]
- **Accuracy:** [Fact-checking requirements, uncertainty handling]
## Output Format
[Show exact structure expected, e.g.:]
Section 1: [Name]
[Description of what goes here]
Section 2: [Name]
[Description]
...
## Quality Checks
Before finalizing your response, verify:
- [ ] **[Criterion 1]:** [Specific, measurable check]
- Test: [How to verify this criterion]
- Fix: [What to do if it fails]
- [ ] **[Criterion 2]:** [Specific check]
- Test: [Verification method]
- Fix: [Correction approach]
- [ ] **[Criterion 3]:** [Specific check]
- Test: [How to verify]
- Fix: [How to correct]
**If any check fails, revise before responding.**
## Examples (Optional)
### Example 1: [Scenario]
**Input:** [Example input]
**Expected Output:**
[Show desired output format and content]
### Example 2: [Different scenario]
**Input:** [Example input]
**Expected Output:**
[Show desired output]
---
## Notes
- [Any additional context, edge cases, or clarifications]
- [Known limitations or assumptions]
Role Selection Guide
Choose role based on desired expertise and tone:
Expert Roles (authoritative, specific knowledge):
- "Senior software architect" → technical design decisions
- "Medical researcher" → scientific accuracy, citations
- "Financial analyst" → quantitative rigor, risk assessment
- "Legal counsel" → compliance, liability considerations
Assistant Roles (helpful, collaborative):
- "Technical writing assistant" → documentation, clarity
- "Research assistant" → information gathering, synthesis
- "Data analyst assistant" → analysis support, visualization
Critic/Reviewer Roles (evaluative, quality-focused):
- "Code reviewer" → find bugs, suggest improvements
- "Editor" → prose quality, clarity, consistency
- "Security auditor" → vulnerability identification
Creator Roles (generative, imaginative):
- "Content strategist" → engaging narratives, messaging
- "Product designer" → user experience, interaction
- "Marketing copywriter" → persuasive, benefit-focused
Key Principle: More specific role = more consistent, domain-appropriate outputs
Task Decomposition Guide
Break complex tasks into 3-7 clear steps:
Pattern 1: Sequential (each step builds on previous)
1. Gather/analyze [input]
2. Identify [patterns/issues]
3. Generate [solutions/options]
4. Evaluate [against criteria]
5. Recommend [best option with rationale]
Use for: Analysis → synthesis → recommendation workflows
Pattern 2: Parallel (independent subtasks)
1. Address [dimension A]
2. Address [dimension B]
3. Address [dimension C]
4. Synthesize [combine A, B, C]
Use for: Multi-faceted problems with separate concerns
Pattern 3: Iterative (refine through cycles)
1. Create initial [draft/solution]
2. Self-critique against [criteria]
3. Revise based on critique
4. Final check and polish
Use for: Quality-critical outputs, creative work
Each step should specify:
- Clear action verb (Analyze, Generate, Evaluate, etc.)
- Expected deliverable (list, table, paragraph, code)
- Success criteria (what "done" looks like)
Common Constraint Patterns
Length Constraints
**Total:** 500-750 words
**Sections:**
- Introduction: 100-150 words
- Body: 300-450 words (3 paragraphs, 100-150 each)
- Conclusion: 100-150 words
Format Constraints
**Structure:** JSON with keys: "summary", "analysis", "recommendations"
**Markdown:** Use ## for main sections, ### for subsections, code blocks for examples
**Lists:** Use bullet points for features, numbered lists for steps
Tone Constraints
**Professional:** Formal language, avoid contractions, third person
**Conversational:** Friendly, use "you", contractions OK, second person
**Technical:** Domain terminology, assume expert audience, precision over accessibility
**Accessible:** Explain jargon, analogies, assume novice audience
Content Constraints
**Must Include:**
- At least 3 specific examples
- Citations for any claims (Author, Year)
- Quantitative data where available
- Actionable takeaways (3-5 items)
**Must Avoid:**
- Speculation without labeling ("I speculate..." or "This is uncertain")
- Personal information (PII)
- Copyrighted material without attribution
- Stereotypes or biased framing
Quality Check Design
Effective quality checks are:
- Specific: Not "Is it good?" but "Does it include 3 examples?"
- Measurable: Can be objectively verified (count, check presence, test condition)
- Actionable: Clear what to do if check fails
- Necessary: Prevents known failure modes
Examples of good quality checks:
- [ ] **Completeness:** All required sections present (Introduction, Body, Conclusion)
- Test: Count sections, check headings
- Fix: Add missing sections with placeholder content
- [ ] **Citation accuracy:** All claims have sources in (Author, Year) format
- Test: Search for factual claims, verify each has citation
- Fix: Add citations or remove/hedge unsupported claims
- [ ] **Length compliance:** Total word count 500-750
- Test: Count words
- Fix: If under 500, expand examples/explanations. If over 750, condense or remove tangents
- [ ] **No hallucination:** All facts can be verified or are hedged with uncertainty
- Test: Identify factual claims, ask "Am I certain of this?"
- Fix: Add "likely", "according to X", or "I don't have current data on this"
- [ ] **Format consistency:** All code examples use ```language syntax```
- Test: Find code blocks, check for language tags
- Fix: Add language tags to all code blocks
Failure Mode Analysis
Common prompt problems and diagnoses:
Problem: Inconsistent outputs
- Diagnosis: Underspecified format or structure
- Fix: Add explicit output template, numbered steps, format examples
Problem: Too short/long
- Diagnosis: No length constraints
- Fix: Add min-max word/character counts per section
Problem: Wrong tone
- Diagnosis: Audience not specified
- Fix: Define target audience, reading level, formality expectations
Problem: Hallucination
- Diagnosis: No uncertainty expression required
- Fix: Add "If uncertain, say so" + fact-checking requirements
Problem: Missing key information
- Diagnosis: Required elements not explicit
- Fix: List "Must include: [element 1], [element 2]..."
Problem: Unsafe/biased content
- Diagnosis: No content restrictions
- Fix: Explicitly prohibit problematic content types, add bias check
Problem: Poor reasoning
- Diagnosis: No intermediate steps required
- Fix: Require chain-of-thought, show work, numbered reasoning
Testing Protocol
1. Baseline test (3 runs):
- Run prompt 3 times with same input
- Measure: Are outputs similar in structure, length, quality?
- Target: >80% consistency
2. Variation test (5 runs with input variations):
- Slightly different inputs (edge cases, different domains)
- Measure: Does prompt generalize or break?
- Target: Consistent quality across variations
3. Failure mode test:
- Intentionally trigger known issues
- Examples: very short input, ambiguous request, edge case
- Measure: Does prompt handle gracefully?
- Target: No crashes, reasonable fallback behavior
4. Consistency metrics:
- Length: Standard deviation < 20% of mean
- Structure: Same sections/format in >90% of outputs
- Quality: Human rating variance < 1 point on 5-point scale
5. Refinement cycle:
- Identify most common failure (appears in >30% of runs)
- Add specific constraint or check to address it
- Retest
- Repeat until quality threshold met
Advanced Patterns
Chain-of-Thought Prompting
Before providing your final answer:
1. Reason through the problem step-by-step
2. Show your thinking process
3. Consider alternative approaches
4. Only then provide your final recommendation
Format:
**Reasoning:**
[Your step-by-step thought process]
**Final Answer:**
[Your conclusion]
Self-Consistency Checking
Generate 3 independent solutions to this problem.
Compare them for consistency.
If they differ significantly, identify why and converge on the most robust answer.
Present your final unified solution.
Constitutional AI Pattern (safety)
After generating your response:
1. Review for potential harms (bias, stereotypes, unsafe advice)
2. If found, revise to be more balanced/safe
3. If uncertainty remains, state "This may not be appropriate because..."
4. Only then provide final output
Few-Shot with Explanation
Here are examples with annotations:
Example 1:
Input: [X]
Output: [Y]
Why this is good: [Annotation explaining quality]
Example 2:
Input: [A]
Output: [B]
Why this is good: [Annotation]
Now apply the same principles to: [actual input]
Domain-Specific Templates
Code Generation
Role: Senior [language] developer
Task:
1. Understand requirements
2. Design solution (explain approach)
3. Implement with error handling
4. Add tests (>80% coverage)
5. Document with examples
Constraints:
- Follow [style guide]
- Handle edge cases: [list]
- Security: No [vulnerabilities]
Quality Checks:
- Compiles/runs without errors
- Tests pass
- Handles all edge cases listed
Content Writing
Role: [Type] writer for [audience]
Task:
1. Hook: Engaging opening
2. Body: 3-5 main points with examples
3. Conclusion: Actionable takeaways
Constraints:
- [Length]
- [Reading level]
- [Tone]
- SEO: Include "[keyword]" naturally
Quality Checks:
- Hook grabs attention in first 2 sentences
- Each main point has concrete example
- Takeaways are actionable (verb-driven)
Data Analysis
Role: Data analyst
Task:
1. Describe data (shape, types, missingness)
2. Explore distributions and relationships
3. Test hypotheses with appropriate statistics
4. Visualize key findings
5. Summarize actionable insights
Constraints:
- Use [tools/libraries]
- Statistical significance: p<0.05
- Visualizations: Clear labels, legends
Quality Checks:
- All analyses justified methodologically
- Visualizations self-explanatory
- Insights tied to business/research questions
Quality Checklist
Before finalizing your engineered prompt:
Structural:
- Role clearly defined with relevant expertise
- Objective is specific and measurable
- Task broken into 3-7 numbered steps
- Each step has clear deliverable
Constraints:
- Output format explicitly specified
- Length requirements stated (if relevant)
- Tone/style defined for target audience
- Content requirements listed (must include/avoid)
Quality:
- 3-5 quality checks included
- Checks are specific and measurable
- Known failure modes addressed
- Self-correction instruction included
Testing:
- Tested 3-5 times for consistency
- Consistency >80% across runs
- Edge cases handled appropriately
- Refined based on failure patterns
Documentation:
- Examples provided (if format is complex)
- Assumptions stated explicitly
- Limitations noted
- File saved as
meta-prompt-engineering.md