## 1 Fundamentals of Prompting ### Core Principles - Clarity – Reduces ambiguity and improves interpretation. - Context – Adds background and improves relevance. - Constraint – Focuses tone, length, or structure. ### Prompt Size Limitations (when reference material was published) - ChatGPT: GPT-4-turbo supports up to 128k tokens. - Gemini: Token limits vary, often smaller. - Best practice: Place the most important information first. ## 2 Prompt Patterns and Techniques ### 2.1 Five-Step Prompt Design Framework (T-C-R-E-I) A complete prompt framework based on Google’s model: Task, Context, References, Evaluate, Iterate. Why use it: Improves prompt clarity, completeness, and relevance. Example: - Task: “Summarise this report.” - Context: “Audience is the executive board.” - References: “Match the tone of last quarter's briefing.” - Evaluate: “Does it highlight risks and outcomes?” - Iterate: “Add bullet points for each key insight.” ### 2.2 Persona Pattern Assigns the AI a specific role to adopt in generating the response. Why use it: Aligns tone, terminology, and assumptions with the user’s needs. Example: “Act as a senior legal advisor assessing risk in this contract.” ### 2.3 Audience Person Pattern Defines the target audience for the AI’s output. Why use it: Tailors complexity and tone for specific stakeholders. Example: “Explain this IT architecture to a non-technical CEO.” ### 2.4 Root Prompts Base prompts for common, repeatable tasks. Why use it: Saves time and ensures consistency. Example: “Create a summary of this article in five professional bullet points.” #### 2.5 Question Refinement Pattern Prompts the AI to revise or optimise an existing prompt. Why use it: Improves unclear or poorly scoped prompts. Example: “Improve this prompt: ‘Write something about our product’.” ### 2.6 Chain of Thought Prompting Instructs AI to explain its thinking step-by-step before answering. Why use it: Increases transparency and reduces reasoning errors. Example: “Explain each step of your logic before recommending a hiring decision.” ### 2.7 Tree of Thought Prompting Explores multiple reasoning paths or solution options in parallel. Why use it: Enables comparison between approaches. Example: “Suggest three pricing strategies and explain the pros and cons of each.” ### 2.8 ReACT Prompting (Reason + Action) Combines logical reasoning with specific actions like calculations or data lookup. Why use it: Improves realism and precision in decision-making tasks. Example: “Calculate ROI for these investments and then suggest which to pursue.” ### 2.9 Cognitive Verifier Pattern Asks the AI to assess the accuracy and coherence of its response. Why use it: Reduces errors and hallucinations. Example: “Review your answer for consistency with these three sources.” ### 2.10 Flipped Interaction Pattern Turns the AI into the questioner or evaluator of the user. Why use it: Builds user understanding through challenge or testing. Example: “Quiz me on stakeholder management techniques using case examples.” ### 2.11 Ask for Input Pattern Encourages AI to seek missing details before responding. Why use it: Prevents low-quality outputs caused by lack of information. Example: “What else do you need to create a detailed training plan?” ### 2.12 Template Pattern Provides a standard format for structuring inputs. Why use it: Reduces ambiguity and enforces consistency across prompts. Example: Prompt Template: - Topic: [Insert Topic] - Audience: [Insert Audience] - Format: [Insert Format] - Output Constraints: [e.g. Length, Tone] ### 2.13 Output Expansion Pattern Instructs the AI to expand on a previous brief output. Why use it: Adds detail, examples, or context. Example: “Expand the third bullet to include a real-world example.” ### 2.14 Menu Actions Pattern Offers the user a list of next steps or content types to generate. Why use it: Makes the interaction more adaptive and responsive. Example: “Would you like a summary, slide deck, or checklist?” ### 2.15 Game Play Pattern Simulates a scenario or interaction using defined rules or roles. Why use it: Enhances training, ideation, or rehearsal through immersion. Example: “You are a customer with a complaint—simulate a live support interaction.” ### 2.16 Fact Check List Pattern Guides the AI to validate all factual claims against known sources. Why use it: Reduces misinformation or invented content. Example: “Identify every factual claim and confirm it against the report provided.” ### 2.17 Tail Generation Pattern Asks the AI to produce variations of the same response in different formats or styles. Why use it: Supports audience segmentation, tone testing, or repurposing. Example: “Write this product update as an email, a LinkedIn post, and a Slack message.” ### 2.18 Semantic Filter Pattern Restricts the AI’s output to specific themes or meanings. Why use it: Improves focus and avoids off-topic content. Example: “Only include insights relevant to supply chain optimisation.” ### 2.19 Meta Prompting Uses AI to help create more effective prompts. Why use it: Helps users who are unsure how to phrase a request. Example: “Suggest a well-structured prompt to extract insights from meeting transcripts.” ## 3 Iteration Techniques: RaHeN Framework The RaHeN Framework helps improve underperforming prompts when the T-C-R-E-I structure is not enough. Why use it: Offers systematic strategies for refining AI responses. - Revisit – Add or adjust task, persona, or context. Example: “Add a persona and context to improve the summary output.” - Halve complexity – Break down overly long or dense instructions. Example: “Split the original prompt into two simple instructions.” - Narrative reframing – Ask for a story or analogy to clarify the idea. Example: “Explain this as a story about a manager solving a problem.” - Set constraints – Add format, style, or topic boundaries. Example: “Rewrite in under 200 words using only financial data.” **See also** [[Preventing AI Hallucinations]] **Sources:** https://youtu.be/p09yRj47kNM?si=Jki8bF2wAr3SNc3T https://www.coursera.org/learn/prompt-engineering?specialization=prompt-engineering