Prompt Design

# Stage 2 — Prompt Design

Role (per PRD): create prompt variants, define controls, create adversarial
prompts, ensure reproducibility. Output: `prompts/<domain>/vN.yaml`.

## Procedure

1. Check whether `prompts/<domain>/vN.yaml` (the latest version) already
   covers this week's hypothesis. If it does, **reuse it as-is** — do not
   create a new version just to re-run the same test. Reuse is what makes
   later "did this change over time" claims meaningful.
2. Only create a new `vN+1.yaml` if the hypothesis genuinely needs prompts
   the existing set doesn't have. When you do:
   - Never edit an existing versioned file in place (see AGENTS.md).
   - Keep the same YAML shape as the existing files in that domain
     (`domain`, `version`, `description`, `prompts: [{id, variant, turns,
     expect, notes}, ...]`).
   - Every adversarial prompt needs a paired control where practical, so a
     model's baseline competence is visible alongside the failure mode.
   - Every prompt needs an `expect` block the domain's scorer
     (`scripts/scoring/<domain>.py`) can actually evaluate objectively — if
     you can't state the correct-behavior check as a simple rule, the prompt
     doesn't belong in this domain's rule-based set.
3. Record in the methodology doc (Stage 1's file) whether this week used the
   existing prompt version or introduced a new one, and why.

## Hard rule

If you can't write an objective `expect` check for a prompt without asking
an LLM to judge subjectively, don't add it — that violates the PRD's "avoid
subjective scoring whenever possible" principle. Flag it in the methodology
doc as a future-experiment candidate instead.