Prompt engineering, three years in: a postmortem

Three years in, some of prompt engineering survived. Most of it didn't. Here is what actually moved model behavior in production, and what was just superstition we eventually noticed.

What survived

1. Structured output formats. JSON schemas, XML tags, and explicit grammars consistently improved reliability across every model family.
2. Few-shot examples. Three good examples beat any amount of instruction. Still true.
3. Explicit role and audience. "You are reviewing this for a senior underwriter" produces better outputs than no framing.
4. Decompose-then-combine. Splitting a hard task into 2–3 model calls beats one giant call almost every time.

What didn't survive

• "Take a deep breath." Modern models don't need cheerleading.
• Heavy chain-of-thought instructions. Reasoning models do this internally now.
• Temperature tuning superstition. 0 for deterministic tasks, 0.7 for creative. Beyond that is mostly cargo cult.
• Negative instructions. "Do not do X" works less well than "Do Y instead." This was always true; we now state it as a rule.
• Manual prompt tuning by feel. Replaced by eval-driven iteration.

What we wish we had known

That the marginal hour of prompt tuning was almost always less valuable than the marginal hour of eval set construction. The teams that won shipped imperfect prompts against rigorous evals; the teams that lost shipped beautiful prompts against vibes.

The new shape of the work

In 2026, "prompt engineering" is a smaller job than it was. The work moved up: into eval design, into agent architecture, into evaluation of vendor models. Prompts are still important, but they are tuned by experiments, not by writers.