Day 3. Advanced Prompt Engineering

When a prompt fails, the fix is almost always a missing component, not a magic word.

Role, task, context, examples, format, audience. Not every prompt needs all six, but the diagnosis of a weak output nearly always points at one of them. A structural engineer who reviews tender documents beats an expert. Write, compare, extract, rank, critique are tasks, help me with is not. Context is what you know that the model does not, and it is where most quality actually comes from.

Say what you want, not only what you do not want. Negative only instruction leaves the model to guess the positive, and it guesses average. Today carries two videos where other days carry three or four, so the reading is primary.

The written lesson. Read it slowly, it saves you later.

The anatomy of a prompt that works

Six components. Role, who is answering, narrow beats grand. Task, the verb, vague verbs produce vague output. Context, the client, the constraint, the thing that went wrong last time. Examples, one or two samples of the output you want. Format, length, structure, what must not appear. Audience, a brief for a CFO and a brief for a site foreman are different documents with the same facts.

Zero shot, one shot, few shot

Zero shot is instruction with no example, fine for common tasks. One shot gives a single example, suddenly the model has a target to match, and this is the single biggest quality jump available for the least effort. Few shot gives three to five examples, now you are teaching a pattern including the edge cases. For anything you will run repeatedly, invest here once.

Chain of thought

Ask for the reasoning before the answer: work through this step by step, then give your conclusion. For arithmetic, logic, planning and multi constraint problems this measurably improves results, because the intermediate steps become part of the context the model conditions on. It is also auditable, you can see where it went wrong instead of only that it did. Newer reasoning tuned models do much of this internally, and asking for visible working is still worth it when you intend to check the answer.

Decomposition beats one giant prompt

A brief with eight requirements will usually see the model honour five. Split it. Outline first, then draft section by section, then a separate pass for tone, then a separate pass for fact checking. Each step is simpler, each is inspectable, and you can stop and correct at any point. This maps directly to how a film runs, you do not shoot, grade, cut and mix in one action.

Prompt as an asset, and debugging

The prompt that worked is a company asset. Save it, version it, note which model and which date, since a prompt tuned to one model can degrade on another. This is the discipline behind every skill file in the GOamplify system.

When output is wrong, resist rewriting the whole prompt. Change one variable and re-run, otherwise you learn nothing about which change helped. Common causes in order: ambiguous task verb, no example, too many requirements in one shot, contradictory constraints, and context that stayed in your head.

Two tracks. Pick yours, produce something.