Prompt Engineering

Placement — instructions buried in the middle of a long context are measurably less likely to be followed than the same words at the start or end. And conflict — an older rule may contradict the new one, and the model resolves that silently rather than flagging it. Why prompts fail →

Providers cache the computed state of a prompt prefix and discount matching tokens steeply, but only while the prefix matches exactly from position zero. Stable content first, volatile content last — and never a timestamp at the top. Anatomy of a prompt →

That the values are true. Structure removes exactly one class of error — the parse error — and converts the rest into well-typed hallucinations. Build honesty into the schema: a nullable answer, a confidence field, an explicit 'insufficient information'. Structured output →

Remove the 'think step by step' scaffolding. The reasoning model was trained to produce its own long reasoning trace, and prescribing your steps can override a strategy better than yours. Manual chain of thought still earns its tokens on small, fast models. Few-shot & chain of thought →

Treating the output as a value when it is a claim. A system that cannot say 'I am not sure' converts every uncertain case into a confident wrong one. The fix is structural: let the model abstain, return a confidence you have actually checked, and route the low-confidence tail to review rather than to the user. Designing for uncertainty →

Not the prompt — the budget. The question stops being 'how do I phrase this' and becomes 'of everything this system knows right now, what earns a seat in the window.' The levers are compaction, tool-result hygiene, externalized memory, and keeping the cacheable prefix stable. Context engineering →