Lesson 4 — Using it well
Two people with identical tools get very different value from them. The difference is rarely prompt wording; it is choosing the right tasks.
The rule that decides everything
Generation pays off when verifying the output is much cheaper than producing it.
Apply this and most decisions become obvious.
Where the rule holds:
| Task | Why it works |
|---|---|
| Drafting text you will edit | Reading and fixing a draft beats staring at a blank page |
| Twenty variations of a headline | Judging which is best takes seconds; writing twenty takes an hour |
| Summarising a document you can check | The source is right there to verify against |
| Extracting fields into a schema | Validation is automatic |
| Explaining unfamiliar code | You can run it and see |
| Translating into a language you read | You will notice if it is wrong |
| Concept art and mood boards | Selection is instant, and nobody ships the output directly |
| Boilerplate code | Tests tell you |
Where the rule fails:
| Task | Why it fails |
|---|---|
| Facts you cannot check | Verification costs as much as research, and the output looks authoritative either way |
| Legal or medical advice acted on directly | Errors are expensive and detection requires the expertise you were substituting |
| Anything requiring current information | The model's knowledge has a cutoff, unless you supply the data |
| Precise arithmetic on many numbers | Checking every figure is as slow as computing them |
| Final client deliverables without review | The failure mode is public |
| Decisions about people | Bias, contestability, and the absence of a reviewable reason |
The pattern is consistent: generation is excellent at producing candidates and poor at being the final authority. Structure the work so a human or a test does the deciding.
Five prompting techniques that reliably help
Most published prompt advice is folklore. These five are worth the effort.
1. Say what the output should look like. Vague requests get vague results. "Summarise this" versus "Summarise this in three bullet points for a non-technical executive, focusing on financial impact" — the second constrains format, audience and emphasis, and those constraints are what makes the output usable.
2. Give examples. Two or three examples of input and desired output convey a format more precisely than a paragraph of description, and this few-shot prompting is the single highest-return technique available. It is particularly effective for anything with a consistent shape: classification, extraction, tone matching.
3. Ask for reasoning before the answer, on hard tasks. "Work through this step by step, then give your conclusion" measurably improves accuracy on multi-step problems, because each intermediate token conditions the next and the model has somewhere to do the work. It also lets you see where the reasoning went wrong. For simple tasks it adds cost and nothing else.
4. Supply the source material. Rather than asking what the model knows, paste in the document and ask about that. This converts an unreliable question into a reliable one, and it is the same principle as retrieval-augmented generation at a smaller scale.
5. Iterate rather than perfecting the first prompt. Get output, say what is wrong, get better output. Three rounds of correction beats one long carefully-engineered prompt in most cases, and it takes less time.
What does not help
Worth knowing so you stop doing it:
- Politeness. Please and thank you change nothing measurable.
- Threats and incentives. "You will be penalised if wrong" and offers of payment are folklore.
- Claiming false urgency. "This is critical for my career" does not improve accuracy.
- Very long role preambles. A brief role statement can help set register. Three paragraphs of persona consumes context and adds nothing.
- Asking the model to check its own work in the same turn. It will often confirm its answer confidently. A fresh conversation with the output pasted in, asked to critique it, works considerably better.
For image generation specifically
Structure the prompt: subject, then setting, then style, then lighting, then framing. "A ceramic teapot, on a weathered wooden table, soft morning light from the left, shallow depth of field, muted palette."
Fix the seed to iterate. Change one element at a time and you learn what each does. Regenerating randomly teaches nothing.
Use negative prompts where the tool offers them, to exclude what keeps appearing uninvited.
Reach for the structural tools when composition matters. Inpainting and ControlNet solve in one step what twenty prompt rewrites will not.
Generate widely, then select narrowly. The cost per image is low enough that producing twenty and keeping one is the efficient strategy.
Keep humans on these
Not caution for its own sake — these are the points where the failure is expensive or invisible:
Judgement about what is worth making. The model has no view on whether the brief is right.
Factual verification. Every name, number, date, citation and quotation, before it reaches anyone.
Anything affecting a person's outcome. Hiring, credit, medical, legal. Bias plus unaccountability plus fluency is a bad combination.
Final review of anything public. The characteristic failure is output that is fluent and subtly wrong, which is exactly what a rushed reviewer approves.
Confidential material. Know whether your inputs are retained or used for training before pasting anything sensitive. This is a procurement question, and where the answer is unacceptable, locally-run models exist for exactly this reason.
In three sentences
The rule that decides what to automate is whether verifying the output is much cheaper than producing it, which is why drafting, variation generation, summarising checkable sources and schema extraction pay off, while unverifiable facts, current information and decisions about people do not. Of the prompting techniques that genuinely help, giving two or three examples returns the most, followed by specifying the output format, supplying the source material rather than relying on the model's knowledge, asking for reasoning on multi-step problems, and iterating instead of perfecting a first prompt — while politeness, threats and long persona preambles do nothing measurable. Keep humans on judgement about what is worth making, on verification of every fact, on anything affecting a person's outcome, and on final review, because the characteristic failure mode is output that reads well and is subtly wrong.