1. Frame the shot before you frame the prompt #
The Matrix opening has a very specific sequence: the green glow, the falling code, the close-up on Trinity's boot, the phone booth. If you don't name each shot, Claude picks its own. So I gave it a step-by-step demand:
You are designing a Lego diorama of the opening scene of The Matrix.
Recreate the following three moments in order:
1. The screen breaking into green falling code.
2. Trinity running in the alley, boots landing on wet pavement.
3. The phone booth with glowing green light.
For each moment, specify:
- Lego pieces needed (exact color and rough quantity)
- Minifigure placement and pose
- Camera angle suggestions for photographing the build
- One "trick" to make the scene feel alive
Notice the "one trick" instruction — that forced Claude to stop listing bricks and start thinking like a display builder. It came back with a great idea: use transparent blue rods for the falling code and a green LED behind the booth.
2. Budget everything #
The first output had 2,500+ pieces, which is absurd. I gave it a second constraint:
Total budget: max 800 pieces. No electric parts. Only standard Lego, no custom third-party.
This is where the LLM really surprised me. It started making trade-offs — trading a large window panel for a stack of trans-green 1x1 plates, cutting the minifigure count to two. That constraint-driven iteration is the difference between a pretty description and a usable build plan.
3. Get a structured, actionable output #
Then I asked Claude to format the final result as a checklist I could actually buy parts from. It ended up with a shot list, a piece count per section, and a suggested build order — logistically, the best day's work I've gotten out of Claude Opus in a while.
The takeaway: don't just ask for a recreation — ask for a production plan. We aren't using Claude for the image; we're using it for the shot design. That's the real prompt engineering win. Give it a budget and a sequence, and it'll design a scene you could assemble with real pieces. I'm halfway through ordering the green parts now.
Next The Race to Beat Cheap AI from China: What It Really Takes →
these real-world AI monetization case studies, with plenty of directly applicable cases.