prompt-alignment-auditlisted

# Prompt Alignment Audit Audit a prompt against the **current, model-version-specific** recommendations of its provider, then produce a citation-backed report and a fully justified rewrite. A rewrite is a hypothesis: prove it with the project's eval harness when one exists; otherwise state plainly that confidence is compliance-based. **Never edit a production prompt file without explicit user approval.** ## Cost policy (non-negotiable) All heavy reading — doc research, prompt-file reading, per-dimension audits — is done by **Sonnet subagents** (`model: "sonnet"`) returning compact structured findings. The orchestrator never reads raw provider docs and never pastes full documents into context. ## Phase 1 — Discover - Bare invocation: scan the project for prompts and model IDs. Grep for model strings (`claude-`, `gpt-`, `gemini-`, `o3`, `o4`), prompt/controller files, template files. Audit everything found. - With an argument: audit only that file or pasted prompt (ask for the target model if not stated and not inferable). - Detect an eval harness: ground-truth files, accuracy runners, `make accuracy`/eval targets. Note it for Phase 5. - Capture the API configuration around each prompt (extended thinking, structured output/tools, caching, temperature) from the calling code — audit agents need it to judge config-dependent rules. - No prompts found → say so and ask for a path. Never guess. ## Phase 2 — Rubric freshness Rubrics live in `rubrics/<provider>.md`,