Mep Ai Led Qa

Guardrails (Mandatory)

1. Read-only mode only.
2. Do not run commands that write to disk or remote APIs.
3. Do not run:
   • mep recipe import-url ... (writes markdown)
   • mep recipe upgrade ... (rewrites markdown)
   • mep sl add-recipe ... (writes desired/plan/apply state)
   • mep sl sync apply ... (mutates Todoist)
4. Return findings to the user in plain language only.

Inputs

• recipes_dir (path containing recipe markdown files)
• Selection mode:
  • explicit slugs/paths, or
  • this_week (query by frontmatter added date)
• qa_depth: quick or release

Workflow

1. Resolve the recipe set from explicit slugs/paths or the date-based this_week selector.
2. Read each recipe markdown and fetch the source page content.
3. Compare title, ingredients, quantities, labels, and method fidelity.
4. Optionally flag shopping-sanity risks without mutating any shopping state.
5. Return a deterministic report with per-recipe findings and an aggregate summary.

Selection Workflow

Explicit list

1. Resolve recipe paths from provided slugs/paths.
2. Skip missing files and report them.

This week

1. Compute week window as Monday->today (local date).
2. Read frontmatter added from each *.md in recipes_dir.
3. Select recipes where added is within the window (inclusive).
4. Sort by added, then filename.

QA Procedure (Per Recipe)

1. Read recipe markdown.
2. Parse:
   • title
   • source URL
   • ## Ingredients pipe lines
   • ## Method steps
3. Fetch and read source page content.
4. Evaluate extraction quality:
   • missing source ingredients
   • hallucinated ingredients
   • quantity accuracy
   • label reasonableness
   • method faithfulness (not rewritten/summarized)
   • title accuracy
5. Optional shopping sanity (still read-only):
   • inspect ingredient lines for obvious duplicate/summing risks
   • flag likely merge/ignore issues as recommendations, not mutations

Parallel Execution

1. Run per-recipe QA in parallel where possible.
2. Use bounded concurrency (recommended 3-5 workers).
3. Keep final report ordering deterministic:
   • added date ascending
   • filename ascending

Output Contract

For each recipe:

• PASS or FAIL
• Severity-tagged findings:
  • critical: missing core ingredients, hallucinations, major quantity errors
  • major: label issues, significant method drift
  • minor: small formatting/consistency issues
• Evidence:
  • source snippet summary
  • markdown line(s) involved

Aggregate summary:

• checked/pass/fail counts
• cross-recipe patterns
• smallest fix surface per issue:
  • prompt: mep-cli/prompts/recipe-extraction.mdxml
  • deterministic shopping logic: mep-cli/src/commands/sl.rs
  • config: $MEP_CONFIG_DIR/cooking.toml

Non-Goals

1. No scripted adjudicator pipeline.
2. No deterministic fallback resolver stack.
3. No automatic edits unless the user explicitly asks for fixes in a follow-up turn.

Acceptance Checks

• Run cargo run --manifest-path ~/development/ai/clai/Cargo.toml -- validate skill mep-ai-led-qa --scope project --project-root ~/development/mise-en-place and expect no [Error] lines.
• Given a recipe source URL and corresponding markdown, when this skill is followed, then the output includes PASS or FAIL, severity-tagged findings, and concrete evidence.
• Given a read-only QA request, when this skill is followed, then it does not invoke mep recipe import-url, mep recipe upgrade, mep sl add-recipe, or mep sl sync apply.

Regression Checks

• Run cargo run --manifest-path ~/development/ai/clai/Cargo.toml -- validate all --scope project --project-root ~/development/mise-en-place and expect this skill to introduce no new hard validation failures.
• Run cargo run --manifest-path ~/development/ai/clai/Cargo.toml -- sync --dry-run --scope project --project-root ~/development/mise-en-place and expect .claude/skills and .config/opencode/skills to mirror .agents/skills.