Key Frameworks

COSTAR is a universal framework for structuring AI prompts, developed by Sheila Teo. Rather than typing a vague request, you systematically define the Context (background and situation), Objective (the specific task), Style (formal, analytical, persuasive), Tone (confident, empathetic, urgent), Audience (who will read the output), and Response format (table, bullets, essay, JSON). Studies show COSTAR-structured prompts improve first-pass output quality by 2.6× compared to unstructured requests, reducing rework and iteration cycles.

The 4D Framework provides a universal model for human-AI collaboration, structuring every AI interaction across four stages. Delegate means identifying which tasks are genuinely high-value for AI augmentation versus those requiring human judgment, relationships, or accountability. Describe means communicating your task with precision using COSTAR and context engineering. Discern means critically evaluating every AI output — never accepting the first response without systematic verification. Deploy means taking full professional responsibility for AI-assisted work, applying data governance and disclosure requirements.

SIFT is a lateral reading framework adapted for AI output verification. When you receive an AI response, Stop before acting on it — resist the pull to accept polished-sounding text at face value. Investigate the source of any claim by checking whether the AI has cited real, verifiable sources. Find better coverage by searching for the claim independently rather than relying solely on the AI's version. Trace any quotes, statistics, or references back to their original context to confirm they haven't been misrepresented. 74% of leaders cite inaccuracy as their top AI risk — SIFT is the systematic antidote.

The Double Diamond is a design thinking framework developed by the UK Design Council, structured as two diverge-converge cycles. The first diamond opens with Discover (broad user research, empathy, and opportunity mapping) and closes with Define (synthesising insights into a sharp problem statement). The second diamond opens with Develop (ideation, prototyping, and creative exploration) and closes with Deliver (testing, iteration, and launch). Applied to AI product design, the framework ensures you solve the right problem before building the right solution — preventing the common failure of building technically impressive AI that nobody needs.

Systems Thinking is a discipline for understanding how components of a system interrelate and how systems work over time. Applied to AI, it reveals why optimising one part of an organisation (e.g., automating a workflow) can degrade the whole (e.g., creating bottlenecks elsewhere or eroding skills). Key concepts include reinforcing loops (where AI success breeds more AI investment, accelerating capability), balancing loops (where resistance or governance slows adoption), and leverage points (the places in a system where a small change produces large effects). The Iceberg Model reminds leaders that visible AI events sit above deeper patterns, structures, and mental models that must be addressed for lasting transformation.

ADKAR is a goal-oriented change management model developed by Prosci, describing the five building blocks required for an individual to successfully change. Applied to AI transformation: Awareness means ensuring people understand why AI adoption is necessary and what happens without it. Desire means addressing the 'what's in it for me' question and building genuine motivation to change. Knowledge means providing the training and frameworks people need to use AI effectively. Ability means giving people practice time, coaching, and safe environments to build real competence. Reinforcement means celebrating wins, measuring progress, and sustaining the change so people don't revert to old behaviours.

Context engineering is the emerging discipline of designing the information environment for AI — structuring all available information (documents, data, instructions, and constraints) to maximise AI performance on a specific task. While prompt engineering focuses on the single message you send, context engineering encompasses the entire information architecture: which documents to include, how to chunk and index them, what system instructions to set, and how to manage the context window. The four principles are Relevance (include only what matters — noise dilutes signal), Recency (prioritise fresh data to bridge knowledge cutoffs), Structure (organise information logically with headers and labels), and Boundaries (define scope explicitly — tell AI what is in and out of scope).

Retrieval-Augmented Generation (RAG) is an AI architecture pattern that grounds language model responses in your organisation's specific, up-to-date knowledge. Rather than relying solely on the model's training data (which has a knowledge cutoff and contains no proprietary information), RAG first retrieves relevant documents from a vector database using semantic search, then augments the model's prompt with those retrieved passages, and finally generates a response grounded in that retrieved context. This dramatically reduces hallucination, keeps responses current, and allows AI to work with confidential internal knowledge without retraining the model.

Cognitive Work Design is an original practical framework for designing AI-enabled work around human capability, not merely task automation. It asks teams to calibrate the level of AI autonomy to the consequence of a task; make objectives, evidence, and system context visible; create active review that a capable person can meaningfully perform; define routes for uncertainty and exceptions; and sustain the professional practice required to question or override the AI. The approach turns the vague idea of keeping a human in the loop into visible decisions about attention, expertise, authority, and workflow design.