Prompt Engineering – How to get the best out of Generative AI tools

On the 16th September, Dr Kitty Hung delivered a thought-provoking and informative presentation focussed on AI Prompt Engineering at the AtkinsRealis Hub in Bristol. During the evening, Kitty covered the following topics, displaying a great level of practical knowledge and insight:

Part 1 – Foundations of Prompt Engineering
Prompt engineering foundations

• Kitty explained prompt engineering as the discipline of obtaining better outputs by treating AI as an instruction-driven system rather than a search engine.
• Effective prompts should use explicit active instructions, targeted questions, relevant context, examples, a defined output format, and factual-accuracy requirements.
• Useful prompting also involves specifying the audience, tone, persona, purpose, scenario, and desired length.

Human expertise and AI validation

• Kitty identified three capabilities needed for effective prompting: technical AI literacy, linguistic skill, and subject-matter expertise.
• Domain expertise remains necessary to validate AI outputs, particularly in areas such as project management, legal work, finance, and banking.
• AI outputs should be checked for accuracy because polished responses can still contain irrelevant or incorrect information. 

Structured and iterative prompting

• Breaking complex work into sequential steps helps manage AI’s working memory and improve coherence and accuracy. 
• Kitty presented chain-of-thought-style interaction as an iterative dialogue: ask the AI to confirm its understanding, propose a plan, review the result, and refine the prompt.  This can provide an audit trail of the Ais work and give assurance about the AI output.
• Meta-prompting (ie: a prompt within a prompt) can make the AI act as a prompt engineer that interviews the user, identifies missing information, and improves the prompt before generating the final output. 

Guardrails, security, and governance

• Kitty recommended safeguards included prohibiting invented information, requiring explicit uncertainty, and requesting citations from reputable sources. 
• Sensitive personal, proprietary, or client information should not be entered into public-facing AI systems; Kitty described abstraction and enterprise solutions as safer approaches (eg: a hypothetical scenario rather than a real scenario). 
• For high-risk systems, API accesses should be tested in a sandbox first before releasing them to the “wild”, with audit trails, token controls, escalation procedures, and human review. 

Part 2 – Agentic AI – from Prompt Engineering to Loop Engineering

Generative AI to agentic and multi-agent systems

• Kitty explained that generative AI mainly responds to prompts, whereas agentic AI can plan, make decisions, use tools, and execute tasks toward a complex objective. 
• Multi-agent systems can divide work among specialised agents such as data collectors, analysts, report writers, and proof readers, coordinated through shared memory or a supervisory agent. 
•  Agent design should define identity, authority, role, permissions, task contracts, communication protocols, memory rules, and escalation thresholds. 

Human oversight and responsible adoption

• Kitty emphasised that humans should retain final decision-making authority, especially for disagreements, high-risk situations, and low-confidence outputs. 
• The presentation raised concerns that excessive reliance on AI could reduce human creativity, empathy, and independent thinking hence leading to humans turning to a “Meatproxy”. 
• The closing message was to use AI to improve communication and productivity while preserving human judgement and values. 

Key Takeaways

• Better AI results depend on precise prompts that define the task, context, audience, examples, format, and constraints. 
• Structured, iterative prompting and meta-prompting can improve completeness, accuracy, and relevance.
• Agentic and multi-agent systems provide greater autonomy but require strong permissions, auditability, security controls, and human oversight.
• AI should support human work without replacing domain expertise, ethical judgement, empathy or creativity.

Paul Johnson
SWWE Network Lead