Skip to content

Prompt Engineering Cheat Sheet

Quick reference guide for crafting effective prompts.

AUTOMAT Framework

Component What to Include Example
Audience Who will use this output? R&D team, IP legal, management
User Persona AI's role/expertise "Act as Senior Polymer Chemist"
Task Specific action required Extract synthesis parameters, format protocol
Output Format and structure Markdown table, JSON, bullet points
Method Approach/methodology ISO standards, systematic review
Assumptions Constraints, boundaries Exclude pre-2020, focus on non-oxide ceramics
Tone Voice and style Technical, formal, conversational

Best for: Functional, structured tasks with clear outputs

CO-STAR Framework

Component What to Include Example
Context Background and situation Market position, project goals, constraints
Objective Goal to achieve Secure funding, convince stakeholders
Style Writing approach Academic, journalistic, executive
Tone Emotional quality Confident, cautious, inspirational
Audience Who will read this VC investors, peer reviewers, management
Response Output format/structure 2-page memo, 5 sections, bullet points

Best for: Narrative, strategic documents requiring rich context

Quick Decision: Which Framework?

Is your task...

FUNCTIONAL & STRUCTURED?          |  NARRATIVE & STRATEGIC?
(data extraction, formatting,     |  (reports, summaries, pitches,
code generation, classification)  |  strategic analysis, proposals)
           ↓                       |              ↓
     USE AUTOMAT                  |        USE CO-STAR

The Red List – Never Share

🚫 Unpublished Research

  • Novel molecular structures
  • Exact synthesis parameters
  • Experimental results (ongoing)
  • Failed experiments (negative data)
  • Grant applications under review

🚫 Commercial Sensitive

  • Exact formulations and ratios
  • Proprietary process conditions
  • Yield data revealing efficiency
  • Cost breakdowns
  • Customer/partner identities
  • Pricing strategies

🚫 Personal & Confidential

  • Employee information
  • Customer data
  • Internal communications with strategy
  • Financial data
  • Legal documents

🚫 Security Sensitive

  • Access credentials
  • System configurations
  • Security protocols
  • Vulnerability assessments

Solution for sensitive work: Use local sandbox (Ollama + Llama)

Hallucination Prevention

Techniques

  1. Explicit constraints: "If data missing, mark 'Not reported'—do not estimate"
  2. Citation requirements: "Cite with DOI for every claim"
  3. Range specification: "Focus only on papers 2020-2024"
  4. Verification instruction: "Flag any uncertainty in your response"

Verification Checklist

  • Citations are real (verify DOIs)
  • Numerical values are plausible
  • Claims align with domain knowledge
  • No internal contradictions
  • Sources match claims

Optimization Quick Wins

1. Batch Queries

❌ Bad: 10 separate queries for 10 papers
✅ Good: 1 query processing all 10 papers

Savings: ~70%

2. Think Before Prompting

❌ Bad: Stream of consciousness, multiple refinement rounds
✅ Good: Plan query with framework, get it right first time

Savings: ~75%

3. Cache & Reuse

❌ Bad: Regenerate monthly literature review from scratch
✅ Good: Cache Month 1, only process new papers in Month 2+

Savings: ~90%

4. Choose Right Model

❌ Bad: GPT-4 for simple keyword extraction
✅ Good: BERT-based NLM for classification/extraction, LLM for reasoning

Savings: ~95% for appropriate tasks

5. Use Templates

❌ Bad: Recreate prompt for each similar task
✅ Good: Template with {VARIABLES}, fill in each time

Savings: ~80% time and consistency improvement

Model Selection Guide

Task Type Best Tool Example
Keyword extraction NLM (BERT) Find all papers mentioning "electrospinning"
Document classification NLM (BERT) Categorize 200 patents by technology
Simple formatting Small LLM (8B) Convert lab notes to template
Literature synthesis Medium LLM (70B) Summarize 20 papers with trend analysis
Strategic analysis Large LLM (GPT-4) Competitive gap analysis and recommendations
Code generation Medium/Large LLM Python script for data analysis

Common Prompt Mistakes

❌ Mistake 1: Vague Task

Bad: "Summarise this paper"
Good: "Extract synthesis methodology with parameter table: temp, pressure, yield"

❌ Mistake 2: No Output Format

Bad: "Extract data from these papers"
Good: "Markdown table with columns: [Author, Year, Method, Key Finding]"

❌ Mistake 3: Missing Constraints

Bad: "What's the melting point of PLA?"
Good: "If melting point reported in paper, extract with page #. If not reported, state 'Not reported'—do not use external data"

❌ Mistake 4: Context Overload

Bad: [500 words of company history for simple task]
Good: [Only task-relevant context, <100 words]

❌ Mistake 5: Wrong Tool

Bad: Using GPT-4 for keyword filtering
Good: Use BERT for extraction, GPT-4 for reasoning

Environmental Impact Reference

Action Tokens CO₂ (g) Water (mL)
Simple query ~100 0.5 10
Complex prompt ~500 2.5 50
Document processing ~2000 10 200
Inefficient workflow (10+ queries) ~10,000 50 1000
Optimized workflow (1-2 queries) ~1,000 5 100

Your goal: >70% reduction through optimization

Emergency Contacts

For technical issues with sandbox: - Check Docker Desktop is running - Restart containers: docker-compose restart - View logs: docker-compose logs

For content questions: - Review relevant section in course materials - Ask in workshop Slack channel - Contact Avgi Stavrou

For data security concerns: - Stop immediately - Report to supervisor - Follow Red List protocol

💾 Download: PDF version of cheat sheet (by Maximiliian Vogel)