Product Update · March 29, 2026

How MovieJam Builds a Perfect Rec Flow in Under 90 Seconds

By MovieJam Crew 5 min read Tags: Product · Automation
Notebooks and sticky notes planning a conversational flow

The WhatsApp experience is deceptively simple—text MovieJam, wait a minute, get three eerily accurate picks. Under the hood we run a tight loop of context capture, scoring, and human spot-checks. Here’s the exact flow we iterate on every week.

Latency target: < 90s Top-3 accuracy goal: 86% Human review coverage: 20%

Step 1 · Warm context intake

Conversation primer

We ask for four data points upfront: mood, companions, available services, and hard no-go genres. If the user skips one, the bot adds a light follow-up (“any services off limits tonight?”) before it starts ranking.

We also track “emotional energy” keywords (tired, wired, need cozy) because they map better to success than genre labels alone.

Step 2 · Scoring + ranking

Recipe

The system drafts 5–7 options, then whittles down to three based on freshness (no repeats in the last 21 days for the same user).

Step 3 · Human-in-the-loop

Editor review

20% of chats route to a human editor (rotating schedule) who can tweak synopses, swap a title, or add snack/playlist notes. Editors tag each change with a reason code so the model learns when we override it.

Step 4 · Delivery + instrumentation

What we watch

67% Sessions that end after the first three picks (no swaps needed).
32s Median time from prompt → first response when no follow-up questions are needed.
11% Chats where we deliberately inject a human editor for surprise & delight.
“The magic trick is balancing personalisation with novelty. We’d rather risk one wild-card pick than send three safe tiles you could’ve scrolled past yourself.” — MovieJam Ops

What’s next

Want the full playbook?

If you’re building a compset concierge or want help tuning your rec system, we can share our prompt templates and QA dashboards.

Request the walkthrough →