write-system-promptlisted

# Write a System Prompt ## When this applies - Writing a prompt that will run more than a handful of times - A model ignores rules that are clearly stated - Outputs are inconsistent between calls - A prompt has grown by accretion and nobody dares edit it ## Do not use for - Data extraction specifically → `extract-structured-data` - Agent tool descriptions → `../../../AI Agent/skills/design-agent-tools` - Building the eval set → `../../../Evaluation and Testing/skills/build-eval-set` ## Inputs to collect first | Input | Why needed | Default if unspecified | |---|---|---| | The task, in one sentence | Everything follows from it | **Blocking** | | Output consumer: code or human? | Decides whether to force structure | Ask | | Failure examples from the current prompt | The most useful input there is | Ask for 3 | | Hard/ambiguous cases | Become the examples | Ask for 2 | | Does it see untrusted content? | Decides delimitation and guardrails | Assume yes if it reads anything external | ## Procedure ### Step 1 — Write the objective and the stop condition first One sentence each. If you cannot write the stop condition, the task is not defined well enough to prompt for. **Stop condition:** both sentences exist and a stranger could apply them. ### Step 2 — List constraints in positive form For each rule, write what the model **should do**, not only what it must not. ``` ✗ Never mention pricing. ✓ When asked about pricing, say you cannot quote prices and offer to connect th