prompt-engineeringlisted

# Prompt engineering Produces the prompt text, not a description of it. **Always show the complete prompt in a marked block.** A summary of a prompt cannot be copied, cannot be diffed, and hides exactly the wording that decides the output. ## When invoked 1. **Pin the target**: which model, what input shape, what output shape, and what counts as a good answer. If the request does not say, state your assumption in one line and continue. 2. **Get a failing example** when improving an existing prompt. One bad output localises the problem faster than any amount of reading. 3. **Pick techniques from the signal table**, not by habit. Every technique added costs tokens on every call. 4. **Write the sections in order** (below), omitting the ones that do not apply. 5. **Check against the list** at the end of this file. 6. **Propose two or three test inputs**, including one that should NOT produce the behaviour — that is what catches an over-broad prompt. ## Section order Include only what applies; keep the order. ``` Role only when a stance changes the answer — "expert X", not "helpful assistant" Context background the model cannot infer Task one specific, actionable command. Strong verb first. Input format what arrives, and how it is delimited Instructions the steps, numbered when order matters Examples omit for zero-shot Output format exact schema or template, with one filled example Constraints short and specifi