prompt-auditlisted

Static sweep of every prompt-assembly and output-parsing site for the five prompt-engineering defects — free-text regex/split/JSON.parse where the provider offers a schema, instructions concatenated with untrusted user or retrieved content in one blob, a structured-output call with no schema validation and no repair path, sampling parameters wrong for a single-answer call (temperature > 0 where the provider exposes it — or a temperature set at all where the provider has removed it and a non-default value is a 400), and a prompt literal with no version id feeding logs + cache key + eval run. Emits one finding per site with <path:line> + a real excerpt + the pattern's closure verb. TRIGGER — any diff touching a prompt, a model call, or an output parser; dispatched by /ai-audit and by @ai-feature-reviewer dimension 2. ANTI-TRIGGERS (do NOT fire) — authoring or improving a prompt (that is /add-ai-feature Phase 4); the prompt-injection exploit judgment or the output→sink review (that is @llm-security-reviewer LLM0
adnanmokhtar/refract · ★ 1 · AI & Automation · score 80

Install: claude install-skill adnanmokhtar/refract

# Skill: prompt-audit ## Premise A production prompt is code: structured, versioned, eval-gated. This skill reads the prompt surface as code and reports where it isn't — five defect classes, each already written as a BAD/GOOD pair in `ai/patterns/prompt-engineering.md`, each anchorable to a line. **Every finding cites `<path:line>` + a real 1-line excerpt from that line + the closure verb that closes it.** For an *absence* finding (no schema validation, no version id) the citation is the concrete site that should have carried it — the parse call, the prompt literal — not "the prompt module". No site, no excerpt, no verb → it is a vibe, not a finding. This skill detects and reports; it does not rewrite the prompt (that is `/add-ai-feature` Phase 4) and it does not grade the injection exploit (that is `@llm-security-reviewer`). ## Adapt to the codebase The five defects are only defects **relative to what the provider offers**. Detect the SDK in use from `_extracted-codebase.md § AI/LLM integration` (or the lockfile) and mirror its mechanism — the audit asks "did this call site use the mechanism this provider has", never "did it use the mechanism I know". | Provider surface | Structured output declared via | System channel expressed as | Sampling controls | |---|---|---|---| | Anthropic SDK | `output_config.format` with `{type: "json_schema", schema: …}` (GA — the older `output_format` is deprecated), and/or tool use: a `tools=` entry with an `input_schema` plus **`strict: