Let the normal prompt fail
Aha: not all ambiguity matters.
Prove: identify the two assumptions that most changed the outcome.
Keep: ambiguity audit
Work With AI / field guide / 8 labs
Prompt Engineering
Stop hunting for magic wording. Make your intent visible enough to steer, compare and improve.
Work With AIPrompt EngineeringField guide
Work With AI field guide
What you'll learn
Design prompts as usable specifications: outcome, situation, material, boundaries, deliverable, interaction mode and acceptance criteria where they matter.
Bring / required
Any AI chat tool and one real task whose result you can judge.
What you'll do
Run a weak baseline first, change one thing at a time, then compare behavior rather than admiring prompt wording.
Take away
An ambiguity audit, reusable briefing pattern, interaction-mode card, prompt experiment record and a clear sense of when prompting is no longer the real problem.
Use the shared Field Guide Lab Method when you run the experiments below: baseline first, change one thing, rerun, measure and keep the evidence.
Start with the misconception
You may be thinking: “There must be a better prompt for this.” Sometimes there is. But prompt engineering becomes useful only when it helps you expose intent, diagnose assumptions and define what a useful result should accomplish.
Aha
Prompt Engineering is not finding magic words. It is designing a usable interface between your intent and the model.
Common traps
- Mega-prompting: more words, no clearer task.
- Persona theatre: “world's best expert” instead of a clear cognitive job.
- Constraint soup: too many rules that conflict or do not materially affect the result.
- Regenerate roulette: retrying until something looks better without knowing what changed.
- Self-certification: treating the model's own critique as independent proof.
- Prompt cargo cult: copying a pattern without understanding the failure it solves.
Look out for
- The prompt keeps getting longer but the important mistake remains.
- You cannot state what “good enough” means.
- The model is missing facts rather than instructions.
- The task requires a tool, action, state or authority that prompting cannot supply.
Progressive labs
CHANGE ONE THING. OBSERVE THE RESULT.
Prompt quality is something you observe in work behavior, not something you judge by how impressive the prompt looks.
Separate purpose from format
Trap: “write 300 words” with no useful objective.
Prove: another person can explain what the output should accomplish.
Keep: outcome + deliverable card
Add only constraints that prevent real defects
Aha: boundaries should protect the outcome, not micromanage every token.
Keep: Five-Part Brief
Use examples as behavioral evidence
Trap: one example becomes an accidental rule.
Prove: the pattern transfers to a materially different input.
Keep: example test
Choose the cognitive job
Answer me · Think with me · Challenge me · Interview me.
Aha: interaction mode often matters more than persona wording.
Keep: interaction-mode card
Diagnose before regenerating
Inspect assumptions, uncertainty and missing information, then revise one defect.
Trap: endless iteration.
Keep: before/after steering record
Separate critique from proof
Aha: self-critique is still model output.
Prove: use source, criteria or external evidence for the claim that matters.
Keep: critique-vs-verify checklist
Run a prompt experiment
Keep task/material/model stable where practical and compare usefulness, defects, rework and acceptance criteria.
Trap: changing five variables at once.
Keep: prompt experiment record
Know when to stop prompt engineering
Prompt work is the wrong lever when the failure is really missing or wrong context, unavailable tools, no durable state, weak verification, unclear authority or a broken workflow.
Next tension
Your instruction can be excellent and the answer can still be wrong because AI does not have the right file, policy, case history or current fact. That is Context Engineering.