AgentSkillsCN

gourmet-research

在repo（gourmet/CITY/{overview,candidates,notes,top-places,excluded}.md）中维护基于证据的美食精选清单：发掘并筛选候选餐厅，通过URL收集多方证据，持续量化评分，发布“必尝榜单”与“备选名单”，记录排除理由以备审计之需，并依据所在地区因地制宜地选择合适的点评与推荐来源（切勿硬性绑定特定国家的平台）。

SKILL.md
--- frontmatter
name: gourmet-research
description: "Maintain evidence-based gourmet shortlists in a repo (gourmet/CITY/{overview,candidates,notes,top-places,excluded}.md): discover and triage candidates, gather multi-source evidence with URLs, score consistently, publish Top Picks/Backups, record exclusions for auditability, and choose appropriate review/guide sources based on the region (don’t hard-code country-specific platforms)."

Gourmet Research

Overview

Build and maintain traceable, comparable, auditable, actionable food recommendations (restaurants/cafes/desserts) with multi-source evidence and consistent scoring.

Use references/repo-spec.md for repo-structure rules and references/templates.md for copy/paste templates.

Workflow

0) Identify the locale + constraints

  • •Confirm the city/region and language context.
  • •Confirm constraints that affect feasibility: opening days, reservation methods, mobility, budget, group size, dietary needs.

1) Discovery (add candidates fast)

  • •Add rows to gourmet/CITY/candidates.md with status: inbox.
  • •Use a maps search link initially if you don’t yet have the exact place link.
  • •Do not delete table rows except for true duplicates or proven incorrect entries.

2) Evidence collection (per place)

  • •Add a detailed evidence section to gourmet/CITY/notes.md (template in references/templates.md).
  • •Prefer 3–5 independent sources that fit the locale (see “Source selection” below).
  • •Update the candidate row with a 1-line summary and status (researching → shortlisted / top / rejected).

3) Score consistently (50-point rubric)

  • •Score each researched place out of 50 with the standard breakdown (in notes.md).
  • •Keep scoring comparable within the same city/region; document uncertainty (use unknown).

4) Publish decisions

  • •top-places.md: Top Picks (high confidence) + Backups (good alternatives) + a short plan/strategy section + a checklist section (names follow the repo’s conventions).
  • •excluded.md: record why a place was rejected (auditable).

5) Verify completion criteria

  • •From within the target gourmet repo, run verify_city.sh CITYDIR to quickly detect “inbox” items, pending decisions, and unchecked research checklist items.
  • •If you’re running the script from outside the target repo, pass --repo-root /path/to/repo.
  • •If your repo enforces required section headers, use verify_city.sh --check-headers CITYDIR.

Source selection (choose what fits the region)

Do not hard-code a single review platform globally. Pick sources that match the locale and availability, and state which you used in notes.md.

Choose sources by asking: “What do locals use, and what has the best coverage for this exact city + price tier?”

Recommended source mix (3–5 total):

  • •Maps + reviews: Google Maps / Apple Maps / local dominant maps app (rating + review count + recurring themes).
  • •Local review platform(s): whatever is locally dominant (examples: 食べログ/Retty in Japan; OpenRice in HK/SG; in the UK often Google/Apple plus strong local editorial coverage).
  • •Local-language review synthesis: summarize recurring patterns (don’t fabricate quotes).
  • •Guide/editorial: reputable guides, local food media, chef interviews, newspapers, Michelin/other guides where relevant.
  • •Official channel: official site/social for hours, closures, reservation method (facts only).

Anti-patterns:

  • •Relying on only one platform.
  • •Using platforms that are irrelevant for the locale if better local alternatives exist.
  • •Treating a single viral social post as evidence.

Resources

  • •references/repo-spec.md: structure + auditability rules (generic).
  • •references/templates.md: doc templates and evidence blocks.
  • •scripts/verify_city.sh: quick structural/completion checks.