🎯 Learning Objectives
- Write meta-prompts that generate other prompts
- Build self-evaluation and scoring frameworks
- Orchestrate multi-step reasoning across complex problems
- Design prompt chains for end-to-end workflows
- Implement error detection and self-correction
- Manage context and memory across long tasks
1. Meta-Prompts — Prompts That Write Prompts
You are an expert prompt engineer.
Generate a high-quality prompt for this task:
"A regulatory manager needs to explain data integrity requirements
to a new IT team in simple, non-technical language."
Include: Role · Audience definition · Tone · Format · Length constraints
Return only the prompt. No explanation.
When to Use Meta-Prompts
- Teaching prompt engineering to colleagues
- Creating templates for repeated use cases
- Generating first-draft prompts for unfamiliar domains
- Building shared prompt libraries for teams
2. Self-Evaluation & Scoring Prompts
Evaluate your previous response against:
1. Clarity (1–5): Is every sentence immediately understandable?
2. Completeness (1–5): Are there obvious gaps?
3. Accuracy risk (1–5): Are any claims likely to be wrong?
4. Usability (1–5): Can this be used without editing?
Score each with a one-line justification.
Then identify the single biggest improvement to make.
3. Prompt Chaining for Complex Workflows
Prompt 1 — Research:
Summarise key regulatory requirements for data integrity
in pharmaceutical manufacturing. Bullet points only.
Prompt 2 — Gap Analysis (uses P1 output):
Based on the requirements above, identify the top 5 compliance
gaps most common in small-to-medium pharma companies.
Prompt 3 — Communication (uses P2 output):
Convert the gap analysis into a plain-language memo
for a plant director. Maximum 200 words.
Benefits of Chaining
- Each step can be optimised independently
- Easier to debug — you can see where quality breaks down
- Outputs can be reviewed and corrected between steps
- Reusable chains become automated workflows
4. Context Compression & Memory Design
- Summarise first — "Summarise this 10-page report in 300 words, preserving all key decisions. Then analyse..."
- Chunking — Process long documents in sections, then synthesise
- State passing — "Given these decisions made so far: [list], now decide..."
⚠️ Context Decay
In long conversations, models "forget" early instructions. Re-state key constraints at the end of long prompts, not just the beginning. Or split into fresh prompts with state re-provided each time.
5. Error Detection & Self-Correction
After generating your response:
1. Re-read each claim. Flag any that could be factually wrong.
2. Check that output matches the requested format exactly.
3. Verify that no instruction was missed.
4. If you find any issue, correct it before presenting the final output.
Format:
[SELF-CHECK PASSED / CORRECTIONS MADE: describe what changed]
[Final Output]
✏️ Module 05 Exercise
Design a 3-step prompt chain for a task in your role: (1) research/gather, (2) analyse/interpret, (3) communicate/deliver. Run all three steps. Add a self-evaluation prompt at step 2 and observe the effect on step 3.
🔑 Key Takeaways
- Meta-prompts multiply your prompt engineering capacity across teams
- Self-evaluation prompts surface quality issues that manual review misses
- Prompt chaining handles complexity better than single mega-prompts
- Context compression is essential for long documents
- Self-correction prompts add a built-in quality gate before delivery
- Reusable chains become the building blocks of AI-powered workflows