# Video Routine: AI-Assisted Family Meal Planning

Use this as the first-pass packet for a Tella-editable video about the family meal-planning PRD. Keep the session replay, human narration, and final Tella edit separate.

## Public-Safe Framing

The public story is the design pattern: a meal-planning assistant with memory. Avoid showing private family names, exact home location, Slack workspace details, recipe account exports, API keys, sheet URLs, or raw household meal history.

## Story Arc

1. Start with the weekly planning friction: choosing meals takes effort, repeats creep in, and recipe collections are hard to use.
2. Establish the system boundaries: Mela stores recipes, Google Sheets stores operational memory, Slack handles review and logging.
3. Walk through Sunday recommendations: recipe tags, meal history, weather, votes, and recent recommendation history produce ten candidates.
4. Walk through daily check-in: Slack collects what was actually eaten and the LLM matches it back to recipes or logs it as unmatched.
5. Close on learning signals: upvotes, downvotes, consumed meals, selected-but-not-eaten meals, ignored recommendations, and repeat avoidance.

## Talk-Track Draft

This starts as a family meal-planning problem, but the useful design move is making the assistant remember outcomes. The recipe collection already lives in Mela, so this system does not try to replace recipe management. It reads exports, enriches recipe metadata, and keeps a separate planning memory in Google Sheets.

Every Sunday, the assistant posts ten meal recommendations in Slack. Each recommendation should explain why it appeared: recent history, family feedback, weather fit, effort level, and variety across cuisine or protein.

The daily prompt is the feedback loop. Asking "What did we eat today?" creates a lightweight history of real meals, not just planned meals. That lets the next recommendation run avoid recent repeats and learn from what the family actually cooked.

One important distinction is planned-but-not-eaten versus recommended-but-ignored. Planned-but-not-eaten can mean the meal is appealing but too much work for the week, missing ingredients, or badly timed. That signal should be tracked separately from a simple downvote.

## Edit Map

| Time | Segment | Editing Notes |
| --- | --- | --- |
| 00:00 | Cold open | State the planning problem without showing private Slack or recipe data. |
| 00:45 | Architecture | Show a simple Mela to Sheets to Slack flow. Keep source names generic if needed. |
| 01:45 | Weekly recommendations | Zoom on the ten-meal output shape and the rationale fields. |
| 03:00 | Daily logging | Show the Slack prompt and matched/unmatched meal result with synthetic examples. |
| 04:15 | Learning loop | Explain votes, consumed meals, ignored meals, and planned-but-not-eaten signals. |
| 05:15 | Close | Name the next implementation step: import recipe exports and create the sheet tabs. |

## Redaction Notes

- Replace real recipe account exports with synthetic or scrubbed examples.
- Blur Slack workspace names, channel names, user names, avatars, and message timestamps if they identify the household.
- Do not show Google Sheet URLs, document IDs, webhook URLs, API keys, weather provider credentials, or home address details.
- Use broad weather examples instead of a precise local forecast unless it is synthetic.
- Avoid exposing raw meal history if it includes travel, health, schedule, or family-specific routines.

## Asset Checklist

- Replay source: pending Codex session ID and JSONL path.
- Screen video: pending silent replay recording or synthetic architecture walkthrough.
- Tella narration: record separately from the replay capture.
- Companion page: `/sessions/meal-planning-routine/`.
- Draft page status: keep `draft: true` until real assets are reviewed.