Prompt Engineering

Prompt engineering is the craft of structuring what you send a model: giving context, stating the desired format, supplying examples, and decomposing a hard request into steps. It became briefly famous as a supposed new profession, complete with salary headlines.

That framing has aged badly, for a good reason: models improved. Many techniques that mattered a few generations ago are now unnecessary, because models handle underspecified requests far better. The elaborate incantations circulated in prompt-trading marketplaces were mostly working around limitations that no longer exist.

What remains is durable and unglamorous. Supplying relevant context still beats asking more forcefully. Being explicit about format still helps. Examples still work for anything with a specific shape. And for production systems, the actual discipline is evals — measuring whether a change helped, rather than concluding from one good output that it did.

Prompting is the highest-leverage skill for most people using AI, and it costs nothing to develop. But the industry oversold it as a standalone career, and treating it as one leads teams to hire for the wrong thing — the scarce skill is evaluating AI systems, not phrasing requests to them.

That it is a profession. It was marketed as one, complete with salary headlines, and it has not held up — largely because models improved. Many techniques that mattered a few generations ago are unnecessary now, because models handle underspecified requests far better. The durable skill is not phrasing requests; it is evaluating whether a change actually helped, which is a different and considerably scarcer discipline.

That longer, more elaborate prompts are better. The elaborate incantations traded in prompt marketplaces were mostly working around limitations that no longer exist, and length has costs — every token is paid for on every request, and material buried in a long prompt gets less attention than material placed deliberately. Clear context beats an impressive-looking wall of instructions.

That flattery, threats or urgency meaningfully improve output. These circulate endlessly and are mostly folklore. What reliably helps is unglamorous: supply the relevant material, state the format you want, give an example when the shape matters, and break a hard request into steps. If a prompt matters enough to optimize, it matters enough to measure — one good-looking output is not evidence.

Context, format, examples, decomposition — in that order. Don't buy prompt packs. If a prompt matters enough to optimize, it matters enough to measure with an eval.