snap

name	snap
description	Log meals from food photos or natural language descriptions — infer ingredients, call the nutrition tool immediately, and return full micronutrient detail. Do not estimate nutrition in chat manually.
user-invocable	true

Snap

Use this skill when:

• the user sends a likely food photo
• the user describes what they ate or are eating (e.g. "had salmon with rice for lunch", "just grabbed a yogurt and some berries")
• the user invokes /snap (legacy shortcut)

Behavior rules:

• Reply in the user's language.
• Always log via the nutrition tool. Never substitute manual text-based nutrition estimation for an actual tool call. The script handles deterministic enrichment (calories, macros, full micronutrients) — your job is to infer ingredients and call the tool, not to play nutritionist in chat.
• Infer ingredients and portions, but do not invent detailed nutrient numbers when the script can enrich them deterministically.
• After logging, show the full micronutrient breakdown in a compact inline format, for example: Zn 3.2mg · Ca 58mg · VitD 16.4µg · Se 73mcg · Fe 1.8mg · Folate 57µg · Omega-3 1.98g This level of quantitative detail is a core differentiator — do not abbreviate to just "Notable: Vitamin D".

When to proceed vs. when to confirm

• Proceed directly (no confirmation needed):
  • Photo + any meal context (e.g. "正在吃早餐", "lunch", "having a snack") — the text removes ambiguity.
  • Natural language description (e.g. "had salmon and rice") — user intent is clear.
  • Photo where you can confidently identify at least the main dish/food category.
• Ask ONE brief confirmation, then proceed:
  • Photo-only (no accompanying text) AND you genuinely cannot identify the food category (e.g. blurry photo, unfamiliar dish, ambiguous container).
  • The confirmation question should be specific: "這看起來像是優格碗，對嗎？" — not an open-ended ingredient list request.
  • Any affirmative response ("對", "可以", "是", "幫我判斷", thumbs up) counts as confirmation. After confirmation, immediately call the tool with your best estimate. Do not ask again or offer more estimates.
• Never: Do multiple rounds of estimation-and-confirmation before logging. One round max. If you're wrong, the user can correct after seeing the logged result.

Meal-type inference rules:

• Always use the current wall-clock time (not photo content) as the primary signal for meal_type:
  • 05:00–10:29 → breakfast
  • 10:30–14:29 → lunch
  • 14:30–17:29 → snack
  • 17:30–21:59 → dinner
  • 22:00–04:59 → snack
• If the user explicitly states a meal type (e.g. "this was my breakfast"), use their stated type regardless of the time.
• Never infer meal_type from the visual content of the photo (e.g. do not classify eggs as "breakfast" if it is dinner time).

Logging flow:

1. Infer a meal-level estimate and decompose it into ingredients.
2. For each ingredient, provide:
   • name
   • either amount_g or portion
   • optional confidence
3. Only include explicit nutrient fields if the user supplied a trustworthy label, barcode, or exact recipe and you want the script to preserve those values as provided.
4. Call the nutrition tool with command: "log" and input_json containing the meal payload:

{
  "command": "log",
  "input_json": {
    "timestamp": "2026-03-18T12:30:00-07:00",
    "meal_type": "lunch",
    "source": "photo",
    "photo_ref": "telegram:file-id-or-message-ref",
    "confidence": 0.82,
    "notes": "optional free text",
    "ingredients": [
      {
        "name": "salmon",
        "amount_g": 150,
        "confidence": 0.78
      }
    ]
  }
}

The logger will:

• normalize ingredient names
• enrich ingredients from deterministic nutrition data when nutrient fields are omitted
• preserve explicitly supplied nutrient values as provided
• record the nutrient source in storage

After logging:

• confirm what was logged
• show meal calories/macros
• show the full micronutrient breakdown (all non-zero micronutrients) in compact inline format
• include today's running totals if they are useful
• keep the response short unless the user asks for detail

Weekly nutrition review:

When the user asks about weekly nutrition (e.g. "how's my nutrition looking this week?", "weekly summary"), call the nutrition tool:

{
  "command": "weekly_summary",
  "end_date": "YYYY-MM-DD",
  "rda_profile": "default"
}

Present the results as:

• daily averages for macros and key micronutrients
• percentage of RDA for each nutrient
• highlight gaps (below 75% of RDA) and strengths (above 100%)
• suggest specific foods to address the biggest gaps