prompt-engineering

Decide what actually goes in the context window — static instructions vs just-in-time retrieval vs conversation history vs tools — with a per-section token budget and lost-in-the-middle-aware ordering. Reach for this when the window is bloating, quality is dropping as context grows, cost/latency is climbing, or you're unsure what to keep vs compress. Pairs with prompt-pattern-selection.

Build the eval/regression set that gates prompt changes — labeled input/expected pairs over the hard cases, a scoring method (exact / schema-valid / rubric / LLM-judge with its caveat), a pass threshold, a CI gate with the model pinned, and injection cases. Reach for this before shipping a prompt, when a tweak silently broke other cases, or to defend against prompt injection. Pairs with structured-output-design.

Choose the prompting pattern — zero-shot, few-shot, chain-of-thought, decomposition/chaining, role framing, or self-consistency — by tracing the task against reliability need and token/latency cost. Reach for this when a prompt is inconsistent, when you're about to add examples 'just in case', or when one prompt is quietly doing several jobs. Pairs with structured-output-design.

Make an LLM return reliably machine-parseable output — choose the enforcement mechanism (native JSON/schema mode, tool/function calling, constrained grammar, or prose+parser), define the schema, and build the parse/validate/repair path. Reach for this when output format drifts, when downstream code parses model output, or when 'return JSON' in prose keeps failing. Pairs with prompt-pattern-selection.