Clawdbites

• ✅ YES: Present the recipe
• ❌ NO (missing instructions or incomplete): Automatically proceed to audio transcription — do NOT stop or ask the user

Completeness check heuristics:

• Has ingredients = contains 3+ quantity+item patterns (e.g., "1 cup flour", "2 lbs chicken")
• Has instructions = contains action verbs (blend, cook, bake, mix, pour, add) + sequence OR numbered steps

Extraction Command

yt-dlp --dump-json "https://www.instagram.com/reel/SHORTCODE/" 2>/dev/null

Key fields from JSON output:

• description — The caption containing the recipe
• uploader — Creator's name
• channel — Creator's handle
• webpage_url — Original URL
• like_count — Popularity indicator

Recipe Parsing

Look for these patterns in the caption:

Macros:

• "X Calories | Xg P | Xg C | Xg F"
• "Macros per serving"
• "Cal/Protein/Carbs/Fat"

Ingredients:

• Lines starting with quantities (1 cup, 2 tbsp, 24oz)
• Lines with measurement units
• Emoji bullet points (🥩 🌽 🧀 etc.)

Sections:

• "For the [component]:"
• "Ingredients:"
• "Instructions:"
• "Directions:"

Output Format

Present extracted recipe cleanly:

## [Recipe Name]
*From @[handle]*

**Macros (per serving):** X cal | Xg P | Xg C | Xg F

### Ingredients
- [ingredient 1]
- [ingredient 2]
...

### Instructions
1. [step 1]
2. [step 2]
...

---
Source: [original URL]

User Actions After Extraction

Let the user decide what to do:

• "Save to my recipes" → Save to Apple Notes (if meal-planner skill available)
• "Add to wishlist" → Save to memory/recipe-wishlist.json
• "Just show me" → Display only, no save
• "Plan this for next week" → Hand off to meal-planner skill

Wishlist Storage

Optional storage for recipes user wants to try later:

memory/recipe-wishlist.json:

{
  "recipes": [
    {
      "name": "Recipe Name",
      "source": "instagram",
      "sourceUrl": "https://instagram.com/reel/...",
      "handle": "@creator",
      "addedDate": "2026-01-26",
      "tried": false,
      "macros": {
        "calories": 585,
        "protein": 56,
        "carbs": 25,
        "fat": 28,
        "servings": 3
      },
      "ingredients": [...],
      "instructions": [...]
    }
  ]
}

Error Handling

If yt-dlp fails:

• Check if URL is valid Instagram reel format
• May be a private account — inform user
• Suggest user paste caption text manually as fallback

If no recipe found in caption (IMPORTANT):

After extracting, scan the caption for recipe indicators:

• Ingredient quantities (numbers + units like oz, cups, tbsp, lbs)
• Recipe sections ("For the...", "Ingredients:", "Instructions:")
• Cooking verbs (bake, cook, sauté, mix, combine)
• Macro information (calories, protein, carbs, fat)

If none found, tell the user clearly:

"I pulled the caption but it doesn't look like the recipe is there — it might just be a teaser or the recipe is only shown in the video itself. Here's what the caption says:

[show caption]

A few options:

1. Check the comments — sometimes creators post recipes there
2. Check their bio link — might lead to the full recipe
3. Describe what you saw in the video and I can help find a similar recipe"

Recipe detection heuristics:

HAS_RECIPE if caption contains:
- 3+ ingredient-like patterns (quantity + food item)
- OR "recipe" + ingredient list
- OR macro breakdown + ingredients
- OR numbered/bulleted instructions

NO_RECIPE if caption is:
- Mostly hashtags
- Just a description/teaser
- Under 100 characters
- No quantities or measurements

Integration with meal-planner

The meal-planner skill can reference this skill:

• When planning meals, check wishlist for untried recipes
• Suggest wishlist recipes that match pantry items
• Mark recipes as "tried" after they're used in a meal plan

Audio Transcription (V2) — MANDATORY FALLBACK

When caption is missing instructions, ALWAYS transcribe the audio automatically. Do not stop and ask the user — just do it. This is the most common case since creators often put ingredients in captions but speak the instructions.

Step 1: Download video

yt-dlp -o "/tmp/reel.mp4" "https://instagram.com/reel/XXX"

Step 2: Extract audio

ffmpeg -i /tmp/reel.mp4 -vn -acodec pcm_s16le -ar 16000 -ac 1 /tmp/reel.wav

Step 3: Transcribe with Whisper

python3 -m whisper /tmp/reel.wav --model base --output_format txt --output_dir /tmp

Step 4: Parse transcript for recipe Look for cooking instructions, ingredients mentioned verbally.

Inference for Missing Measurements

ALWAYS infer quantities when not provided. Never present a recipe without amounts — estimate based on context and standard package sizes.

Vague Language → Specific Amounts

Standard Package Sizes (when item mentioned without amount)

Context-Aware Scaling

By recipe type:

• Stir fry for 2 → 1 lb protein, 4 cups veggies
• Soup/stew → 1.5-2 lbs protein, 4 cups broth
• Sheet pan meal → 1.5 lbs protein, 3-4 cups veggies
• Appetizers → smaller portions, estimate ~12-15 pieces per batch

By servings mentioned:

• "Serves 4" → Scale standard amounts for 4
• "Meal prep for the week" → Assume 5-8 servings
• No servings mentioned → Default to 4 servings

By protein target (if user has macro goals):

• 40-50g protein per serving → ~6-8oz cooked meat per portion
• Scale recipe protein accordingly

Output Format

Always present inferred amounts clearly:

### Ingredients
- 1 lb ground turkey *(estimated)*
- 1 medium onion, diced *(estimated)*
- 2 cups broth *(estimated based on typical soup)*

Mark inferred quantities with (estimated) so user knows what came from the source vs inference.

Combined Extraction Flow

1. TRY CAPTION (instant)
   └── yt-dlp --dump-json → parse description
   └── Recipe found? → DONE ✅
   └── Check for "pinned" / "in comments" / "check comments" → FLAG
   
2. IF FLAGGED: CHECK FOR CREATOR COMMENT
   └── Look through comments for creator's username
   └── If creator comment found with recipe → DONE ✅
   └── If not found → continue + notify user

3. TRY AUDIO (30-60 sec)
   └── Download video
   └── Extract audio with ffmpeg
   └── Transcribe with Whisper (base model)
   └── Parse transcript for recipe
   └── Infer missing measurements
   └── Recipe found? → DONE ✅

4. PRESENT RESULTS + PROMPT IF NEEDED
   └── Show what was extracted from audio
   └── If "pinned" was flagged, tell user:
       "The creator mentioned the full recipe is pinned in the comments.
        I extracted what I could from the audio, but if you want the 
        exact measurements, paste the pinned comment here and I'll 
        merge it with what I found."
   
5. TRY FRAME ANALYSIS (if audio incomplete)
   └── Extract 5-8 key frames with ffmpeg
   └── Send to Claude vision
   └── Ask: "Extract any recipe text, ingredients, or measurements shown"
   └── Merge findings with audio transcript
   
6. FALLBACK (nothing found)
   └── Inform user: "Recipe wasn't in caption or audio/video"
   └── Offer: search for similar recipe based on video title/description

Frame Analysis

Extract key frames and analyze with vision model.

Extract frames:

# Extract 1 frame every 5 seconds
ffmpeg -i /tmp/reel.mp4 -vf "fps=1/5" /tmp/frame_%02d.jpg

# Or extract specific number of frames evenly distributed
ffmpeg -i /tmp/reel.mp4 -vf "select='not(mod(n,30))'" -vsync vfr /tmp/frame_%02d.jpg

Send to vision model: Use Claude's image analysis to read each frame:

• Recipe cards / title screens
• Ingredient lists shown on screen
• Measurements in text overlays
• Step-by-step instructions displayed

Vision prompt:

Analyze this frame from a cooking video. Extract any:
- Recipe name or title
- Ingredients with quantities
- Cooking instructions
- Nutritional information / macros
- Any other recipe-related text shown

If no recipe text is visible, respond with "No recipe text found."

Merge strategy:

• Audio transcript = primary source (spoken instructions)
• Frame analysis = supplement (exact measurements, recipe cards)
• Combine both, prefer specific measurements from visual over inferred from audio

Pinned Comment Detection

Scan caption for these phrases (case-insensitive):

• "recipe pinned"
• "pinned in comments"
• "check comments"
• "in the comments"
• "comment below"
• "recipe below"
• "full recipe in comments"

If detected, flag and notify user after extraction:

"Heads up — the creator said the recipe is pinned in the comments. I got what I could from the audio, but yt-dlp can't access pinned comments without login. If you want the exact recipe, copy the pinned comment and send it to me — I'll format it properly."

Requirements

• yt-dlp — brew install yt-dlp
• ffmpeg — brew install ffmpeg
• whisper — pip3 install openai-whisper (runs locally, no API key)
• No Instagram login required for public reels