Gourmet Research

Template-First Workflow (Summary)

1. Initialize: Ask for output language + city, then copy templates from assets/templates/ into the city folder.
2. Normalize: Update language/city placeholders in the copied files before research begins.
3. Discovery: Capture raw ideas in inbox.md, then move top candidates into candidates.md with status: inbox.
4. Evidence: For each candidate, write a full evidence block in notes.md with sources + practical constraints.
5. Score: Apply the 50-point rubric and justify each component in notes.md.
6. Decide: Promote to Top Picks (>=35), Backups (30-34), or reject (<30).
7. Publish: Update top-places.md and excluded.md to match decisions.
8. Verify: Ensure no inbox statuses remain and required sections exist.

Ranking Retrieval (When user asks for “highest score”)

Before extracting any “top N” list, confirm the scope:

• Geography: Okinawa prefecture vs main island only vs specific subarea.
• Category: overall vs cuisine category.
• Source URL: must match the user’s intent exactly.

Checklist (must pass):

1. URL matches the requested scope (prefecture vs category).
2. If “main island only” is required, exclude island subareas (A4705/A4706).
3. Page title confirms the intended ranking.
4. Language modal handled so list items actually render.

If static scraping fails or content is blocked, use Playwright to load the page, close the language modal (日本語), and then extract items.

Evidence & Negative Review Rules

• Sources must include: Maps + local reviews + guide/editorial + official channel (where available).
• Negative review analysis is conditional: perform a focused negative review pass when risk signals appear in any source.
  • Risk signals: repeated service complaints, hygiene/safety concerns, tourist-trap claims, extreme queue issues, inconsistent ratings, unclear hours/reservations.
  • If triggered: add a Negative reviews subsection in notes.md, adjust Risk/Consistency/Value as needed, and sync scores/status across files.

Locale-Specific Source Suggestions (Optional)

Scoring (50-Point Rubric)

• Taste/Quality (0-10)
• Value (0-10)
• Convenience (0-10)
• Consistency (0-10)
• Risk (0-10, higher = lower risk)

Thresholds:

• Top Picks: >=35
• Backups: 30-34
• Reject: <30 (or hard exclusion: hygiene/safety/tourist-trap evidence)

Roles (Optional, Compact)

• Research: find sources + capture evidence.
• Verify: resolve conflicts, confirm practical constraints.
• Score: apply rubric + justify.
• Synthesize: finalize top-places + dining strategy.

Quick Reference

Example (Evidence Block)

### Sakura Teahouse
**Official**: https://example.com
**Maps**: 4.4/5 (820 reviews) - https://maps.app.goo.gl/...
**Local reviews**: 3.7/5 (420 reviews) - https://tabelog.com/...
**Guide/editorial**: https://guide.example.com/...
**Notes**: quiet seating, popular seasonal desserts
**Practical**: reservations recommended, closed Tue
**Score**: Taste 8 / Value 7 / Convenience 6 / Consistency 7 / Risk 7 = **35/50**

Common Mistakes

• Skipping templates and mixing content across files.
• Skipping inbox.md and dumping raw ideas into candidates.
• Translating place names instead of using the original language.
• Using only one review platform.
• Pulling the wrong ranking scope (category vs overall, islands included).
• Changing scores without updating candidates/top-places/excluded.
• Ignoring unclear hours or reservation policies.

Rationalization Table

Red Flags — Stop and Fix

• Candidates deleted instead of rejected.
• Scores updated in notes but not in candidates/top-places.
• Missing output language decision.
• Uncited claims or ratings.

References

• references/repo-spec.md
• assets/templates/