My AI built me a teleprompter 10 min read

My hobby project

My AI built me a teleprompter.

I filmed a whole video through it. This page shows what the AI did, what I did, and why it matters to people who will never touch a 3D printer.

10-minute read · No engineering knowledge needed

Florian standing next to the black 3D-printed teleprompter on its tripod.
Florian Roscheck Manager AI Consulting @ Henkel Hobby project · own time, own views
  • 16days from first brief to filming
  • 26printed parts, all designed by AI
  • 20,000+lines of code, written by AI
  • 1human, making the calls
39 seconds, filmed through the device itself.

The short version

A device that didn’t exist, described in words

I described what I needed. An AI called Claude designed the device, tested it on the computer, wrote the files for my 3D printer, and worked out why things broke.

I set the goal, made the big decisions, printed the parts and put them together.

The gadget is fun. The real story is how many different jobs one AI could do in a single project.

The device

What is a teleprompter?

It lets you read a script while looking straight into the camera. The trick is a sheet of special glass.

The script reflected in the glass, readable from where the speaker stands.
From the speaker’s side, the script floats in the glass.

One sheet of glass: a mirror for you, a window for the camera.

Mine had to work with three things I already owned: a pocket-sized DJI camera, an old Lenovo tablet and a sheet of teleprompter glass. And it had to fold to the size of a book, tablet and glass inside.

I couldn’t find another one like it.

Florian holding the folded teleprompter, which is about the size of a thick book.
Folded: 239 × 159 × 43 mm. Tablet, glass and hood are inside.

Step 1 · The brief

First, write down what “good” means

Before any design, there was a written brief: about 4,000 words with 14 hard rules, agreed in an interview. A few of them, in plain language:

  • Folds to 240 × 160 × 45 mm, tablet and glass inside
  • The glass never comes out, not even for folding
  • No stray light may reach the lens
  • The camera’s view is never cut off
  • Every custom part prints on my home 3D printer
  • At most five test prints
  • Stop twice and wait for my approval
  • Never call a test passed unless it really ran

The last two rules turn a fast assistant into a colleague you can trust.

Step 2 · The first surprise

The AI’s first move was to disagree with me

My brief said the glass should sit at about 45 degrees, the classic teleprompter angle. Claude did the maths before designing anything and flagged a conflict: with my glass and my camera, 45 degrees cannot work.

Try it yourself, with results from the project’s own optics code.

45°My brief

Script reaching the speakerCamera’s viewPart of the view misses the glass

Share of the tablet’s screen the speaker can see in the glass

53° is the sweet spot. Claude’s first answer was 59°. With better lens data, which I checked by measuring the camera’s view on a wall, it moved to 53°. The visible script area roughly doubled.

It never quietly changed my number. It logged the conflict and waited for my OK.

Step 3 · Three ideas

Three designs, then one human decision

Claude built three different designs as real 3D models and scored each one on strength, weight, ease of packing, printability and repair.

The Tower design, set up.
Set up
The Tower design, folded.
Folded

One detail is worth noticing. The Capsule won on points, 3.15 to 3.00. Claude recommended the Tower anyway and explained why: the lead came mainly from lower weight, while the Tower had no sliding parts to fine-tune.

The score was advice, not a verdict. I chose the Tower.

Step 4 · The design

It doesn’t draw. It writes.

When I studied engineering, I built parts by clicking through a 3D design program for hours. Claude works differently. It writes the design as code: a long, exact recipe that says “take this plate, cut a hole here, add a hinge there”.

Then · 2011 · Me, clicking
A bottle gripper Florian designed in a classic 3D design program in 2011.
Now · Claude, writing
# printed floor under the nut: a 4.5-mm tripod
# screw still engages 2.5 mm = 2 turns (F-13)
NUT_FLOOR = 2.0

# F-56 (P14): gusset under the knuckle
y_gus = sorted([Hn_y - s * FRAME_KNUCKLE_R, …])
lug_n = lug_n + (GU.box(xs[0], xs[1], *y_gus, …)
Real lines from the design file.
Exploded view of the teleprompter, showing how the printed parts fit together.
Every printed part comes out of that recipe.

Why a recipe beats a drawing:

  • Easy to change. Alter one number, run it again, and everything follows.
  • On record. Each number sits next to its reason.
  • Testable. A computer can check a recipe automatically, over and over.

Step 5 · Testing

It tried to break the design on the computer first

You can’t buy a teleprompter simulator. So Claude wrote its own test tools and treated the 3D model as a digital twin.

› Ready to film
0 of 779 positions checked0 collisions
Rendering of the camera's field of view passing through the glass.
Camera view0 rays clipped, in both video formats.
Rendering of test points inside the hood, each checked for a path to the outside.
Light leaks299 points tested, 0 leaks.
Studio rendering of the folded teleprompter with its lid closed.
Strength and heatEstimates, such as lid sag onto the glass.

How do you test for light leaks without a real device?

Light travels in straight lines. If no straight line gets from outside to the lens, no light does either. Move the flashlight and compare the two joints.

Claude ran this logic in 3D. Such checks prove the geometry, not that a part survives real life.

Step 6 · Reality

Then reality pushed back

A simulation only knows what it has been told. So I printed small test pieces, measured them and typed in the results. Claude explained each failure and reworked the design.

A pile of small 3D-printed test pieces in black and white on a desk.
Test pieces: only the risky bits of the design.

What each test print showed

Checks passedChecks failed

I later raised my limit of five test prints, step by step, to ten.

56 problems were logged. Where did they show up?

The last ten were the hard ones. Here are three.

Problem 1

The tablet was bigger than its file

My 3D file said 210 mm, my calipers 213 mm. I had told Claude to trust the file, so one part no longer fitted. A redesigned lid now holds the tablet.

Lesson: AI can’t check what it can’t see. Wrong input, wrong output.

Problem 2

The hinges snapped

Both small hinges on the glass frame broke off. I sent Claude a photo. It traced the cause to a 4.6 × 2.4 mm neck sitting on one print layer. 3D prints are stacks of thin layers and, like paper, they split between sheets.

Claude called it a design error: a situation it had never checked. The fix adds a foot and a wedge, by estimate about 9.5 times stronger.

Lesson: the most useful words from a colleague are “my mistake, here is why”.

A snapped hinge piece lying next to the black glass frame.
The real break
New hinge with a green foot and wedge. Old hinge with the thin neck marked in red. BeforeAfter

Problem 3

It called off its own fix

The first hinge fix needed a small cut in a neighbouring part. I approved it and started printing. Then Claude’s light check found the cut would let light in. It withdrew the file and I stopped the printer. A better fix followed the same day.

Lesson: hard rules work. “No stray light” beat a plan I had already approved.

The withdrawn print file for the glass frame, stamped withdrawn. Withdrawn
Claude

› Light path through the deck cut.

⚠ Withdrawing this print file.

Who did what

So who actually did the work?

The honest split, stage by stage:

The brief

Set the goal and the hard rules.

Turned the rules into checks. Flagged the 45° conflict.

Concepts

Chose the Tower. Approved 53°.

Built and scored three designs. Reworked the chosen one.

Detailed design

Measured the camera. Made the calls when options came up.

Wrote all 26 parts as code. Wrote its own test tools.

Test prints

Printed, measured, reported what I saw.

Picked what to test, wrote the printer files, evaluated every result.

Buying and building

Paid. Ran 50+ hours of printing. Assembled it and filed plastic.

Checked screws against the design. Wrote the manual. Diagnosed each failure.

The video

Stood in front of the camera. Reviewed every cut.

Script, renders, graphics, sound, edit, captions.

100 logged decisions. Tap a square.

Each square is one entry in the project’s decision log, in order.

Claude made most of the technical calls. Every approval, every extra print and every accepted risk was mine.

What steering looked like

yes to the second coupon; go with geometric overlaps

go with the +1.5 mm panel

my decision is to move ahead, even without a test print

  • ≈ 2,100words typed by me in about 70 messages
  • 2,500+actions by Claude: commands, file edits, checks

Since 12 Oct, without the brief and measurement sheets. Earlier logs are on my old laptop.

The real point

One AI, a dozen jobs

Many people still think of AI as a chatbot that writes text. In this single project, the same AI took on tasks from a dozen different professions.

    Why can one AI do all of that?

    Because modern AI no longer just answers. It uses tools. It writes and runs code, reads files, searches the web and operates programs. When a tool is missing, it writes one.

    YouGoal and rules
    ClaudePlans and acts
    Tools
    • Code
    • Simulation
    • Print software
    • Web
    YouCheck and decide

    If you can describe a job in words and check the result, AI can probably help with it.

    It wasn’t equally good at everything, and it doesn’t replace an experienced engineer. But the range surprised me.

    Your turn

    You don’t need a 3D printer for this

    It comes down to three habits from my day job:

    1. Requirements

      Say what good looks like, and what must never happen.

    2. Execution

      Let AI work in steps, with options, evidence and a log. Decide where it stops for you.

    3. Feedback

      Bring reality back in. Say what went wrong and let it find out why.

    What could that look like in your world?

    Try it · a starting prompt
    I want to work on: [your project].
Good looks like: [3 to 5 measurable points].
Must never happen: [your hard limits].

First, ask what you need to know.
Then give me three options, recommend one and wait for my decision.
Mark anything you have not verified.

    At work, use the AI tools your company has approved, and follow its rules for confidential data.

    The fine print

    What has not been proven

    This is a one-off hobby build, not a product. It works well enough to film with, but these tests have not been run:

    • Flashlight test for light leaks on the real device
    • 1 to 2 hours of continuous filming (heat)
    • A drop from 30 to 50 cm
    • Pressure inside a backpack
    • 250 fold cycles
    • Final fit check of the reinforced hinges

    The AI made real mistakes too. One of its own checks wrongly reported a clear path for a steel pin. That is why the logs and human gates exist.

    My hobby project

    What would you build?

    The design, the code, the video and this page all came out of conversations with an AI. Maybe your next project can too.

    Studio rendering of the finished teleprompter.