recipe-scout

Source-policy + normalization layer for Chinese recipes. Not a single-site scraper.

Source Ranking (strict priority)

Tier A — Structured recipe pages (primary)

• 下厨房 (xiachufang.com) — largest Chinese UGC recipe platform, step-by-step photos, strong search
• 美食天下 (meishichina.com) — broad category coverage, reliable ingredient lists
• 豆果美食 (douguo.com) — large recipe app ecosystem
• HowToCook (github.com/Anduin2017/HowToCook) — programmer-authored, precise gram-level measurements, no fluff; raw URLs: https://raw.githubusercontent.com/Anduin2017/HowToCook/master/dishes/<category>/<dish>/<dish>.md; categories: meat_dish, aquatic, vegetable_dish, staple, soup, dessert, condiment
• Reputable cooking blogs with full recipe text

Tier B — Video + transcript (secondary)

• YouTube / Bilibili cooking videos with captions or clear on-screen ingredients

Tier C — Social posts (fallback / inspiration only)

• 小红书, short-post platforms, comment threads
• Mark confidence low unless validated by another source
• Do NOT treat as authoritative recipe specs

Anti-pattern: Do not make 小红书 the primary recipe backend. UI changes often, ingredient/timing details are frequently missing, extraction is brittle.

Retrieval Strategy

1. Parse dish intent + constraints (cuisine, diet, equipment, time, restrictions)
2. Search Tier A first with Chinese queries (see references/query-examples.md)
3. Collect 3–7 candidates
4. Deduplicate by core technique + ingredient profile
5. Rank by: completeness → clarity → home-cooking fit → ingredient accessibility
6. Use Tier B/C only to fill gaps or add variants

Normalization Rules

• Language: Chinese default (English if user requests)
• Units: metric preferred (g / ml / tbsp / tsp); preserve source units if conversion uncertain
• Times: separate prep / cook / total; mark missing estimates as estimated: true
• Seasoning: keep exact source values; record 适量 as-is with a practical note
• Heat: normalize to 小火 / 中小火 / 中火 / 中大火 / 大火
• Safety: include basic food safety notes (chicken/pork doneness, etc.) when relevant

Anti-Hallucination

• Never fabricate ingredient quantities, temperatures, or timings
• Missing fields → unknown / 未注明
• Clearly distinguish: source facts vs inferred estimates vs your recommendations

Output Modes

Schema: See references/schema.md Obsidian template: See references/obsidian-template.md

Execution Playbook

When invoked:

1. Restate target dish + constraints in one line
2. Search Tier A with Chinese queries
3. Return candidate table (3–7): source | completeness | style | confidence
4. Pick one "default cooking path"
5. Normalize into canonical schema
6. Output quick summary in chat
7. If export requested → write Obsidian note(s), return file paths

Browser vs Web Tools

• Prefer: web_search (Brave Search API) + web_fetch for structured recipe pages — this is the primary retrieval method
• Brave Search tip: Use Chinese queries for Tier A sources (下厨房, 豆果, 美食天下); English queries for fallback/Tier B
• Use browser only when: page requires JS rendering, login/session, or video transcript interaction
• If Chrome relay is unstable → switch to openclaw browser profile

Hard No's

• No aggressive scraping
• No paywall/login bypass
• No social-media hearsay presented as precise recipe specs
• No invented measurements