Prompting is interface design: context, constraints, examples — then verify
Prompting, model behaviour, and human–AI interaction design
The question: How do we turn a general model into a reliable collaborator for a specific task?
Prompting is the fastest way to prototype AI value — and the first place business users learn the discipline of context, constraints, examples, and verification. Done well, prompts mature into reusable task protocols: organisational assets, not one-off magic words.
What the lesson covers
A prompt is a task specification. The reliable ones specify: the task, the context (who, what data, what situation), constraints (what to avoid, what format), examples (few-shot: show 1–3 good outputs), the audience, and the evaluation criteria. Model behaviour is shaped by instruction order, examples, tool access, and randomness (temperature) — so the same intent phrased loosely produces scattered results.
Core patterns: zero-shot (just ask), few-shot (show examples), role prompting (set expertise and perspective), decomposition (break hard tasks into steps), structured outputs (demand JSON/tables for anything downstream), critique loops (generate → critique → revise), and chain-of-verification (make the model list claims, then verify each). "From prompt engineering to context engineering": the frontier is what information you give the model — instructions, examples, tools, data, memory — not incantations.
Good human–AI workflows separate four jobs: generation (the model drafts), critique (the model or a human challenges), verification (a human or a tool checks facts against sources), and decision (a named human owns it). Collapse these into one step and you get automation bias — fluent wrongness accepted at face value.
Prompts should be managed like code: versioned, tested against a small eval set, and shared as team protocols/SOPs. A "prompt library" with test cases and failure-mode notes turns individual skill into organisational capability — and it is the cheapest AI asset you can build this quarter.
Know the failure modes: prompts that assume the model knows your context (it doesn't), untestable asks ("make it better"), hidden assumptions, no output contract, and no verification step. The fix is always the same: specify, exemplify, constrain, verify.
Key points
- A prompt is a task spec: task, context, constraints, examples, audience, evaluation criteria.
- Patterns that matter: few-shot examples, role, decomposition, structured output, critique loops, chain-of-verification.
- Separate generation / critique / verification / decision — collapsing them creates automation bias.
- Prompts are assets: version them, test them, share them as team protocols.
- Score outputs with rubrics, not vibes.
Framework — CLEAR Prompt Protocol
Context (situation + data), Logic (steps to follow), Examples (1–3 model outputs), Audience (who reads it), Requirements (format, length, constraints, what to refuse). Run every recurring prompt through CLEAR and store it with test cases.
The lab
Run a prompt tournament: turn a weak prompt into a reusable task protocol and prove it is better.
Open this lesson, its lab and its quiz
Sources and further reading
- Prompt engineering guide — Anthropic
- Prompt engineering guide — OpenAI
- Co-Intelligence — Ethan Mollick