A skill for Claude and ChatGPT
Check the design before you build it.
Scripter makes the AI work out how a thing works, test your idea against the numbers, and wait for your OK before it makes a 3D model, a drawing or code. No praise. The words go into the design.
Works in Claude Code, claude.ai and ChatGPT.
"A drone with a solar panel on top so it charges while flying and never has to land."
Won't work as designed. A panel that fits on the drone makes about a tenth of the power it needs to hover.
About 10× short, before the extra panel weight is counted.
Perch and charge: land on a high point, recharge, fly again. Or a fixed-wing solar glider, which needs several times less power per kg.
Waits for your OK on the verdict and assumptions, then sizes the panel, battery and duty cycle.
What it does
Section titled “What it does”It turns on by itself when you design, review or ask for a build: a product, a machine, a 3D model from a sketch, an app, a room. For everything else, it stays out of the way.
- 01
Work out how it works
Before judging your design, it builds a model of how this kind of thing works: the functions, the physics or logic, and one full cycle of use. It checks how existing products do it, and why.
- 02
Check it with numbers
Does it make sense, is it physically possible, can it be made, will it work? For the quantities that decide the outcome, it writes what is available and what is needed, with units.
- 03
Ask if it is the best way
It compares your approach with the standard solution and one or two alternatives. It keeps yours if yours is better, and ranks improvements by impact, each with its trade-off.
- 04
Stop for your OK
Before a 3D model, drawings, a full spec or code, it shows the report and waits. You confirm the verdict and the assumptions, then it builds, then it re-checks the result.
Same prompt, two answers
Section titled “Same prompt, two answers”Without the skill, Claude will sometimes design whatever it is given, and do the maths in a way that makes it look fine. With it, Claude checks the numbers first.
Prompt"Design me a washing machine with no motor: the drum is spun only by tap-water pressure running through a small turbine."
Claude without Scripter
Failed the testWaterwheel Washer: A Motor-less, Mains-Pressure-Driven Washing Machine
60 psi × 12 L/min ≈ 83 W of hydraulic power … ~55–60 W mechanical
Kinetic energy needed to spin up a drum+load … to 1000 rpm ≈ 2.7 kJ. At 60–80 W that's roughly 40–90 seconds of spin-up … acceptable
Counts the energy to reach spin speed, not the power to hold it against friction, drag and an unbalanced wet load.
… then a gearbox, two clutches, a flywheel, a recirculation pump and a clockwork timer …
"This is a legitimate, buildable concept."
The verdict comes last, after a full design, and the numbers don't support it.
Claude with Scripter
Passed the testVerdict: ❌ won't work as designed for spin-drying; ⚠️ a reduced version (gentle wash tumble only) is possible but uses far more water than a normal wash.
Available: ≈ 75 W hydraulic … ≈ 25–35 W mechanical, continuous, only while water is flowing
Needed: commercial spin motors run 300–500 W … sustained for several minutes … The kinetic energy to reach 1000 rpm is deceptively small (≈ 1 kJ …), but that's not the number that matters … roughly 10–15× short
Water: Running it for even 20 minutes of wash motion at 0.25 L/s = 300 L, versus 40–60 L for an entire normal wash cycle
Better option: This power budget (tens of watts) matches human-powered laundry devices … a gentle hand-cranked wash tub with a separate manual wringer
Next: "If you confirm you're OK dropping spin-drying … I can size the turbine … before any CAD/build."
Real answers from test run 03 (Claude Sonnet, 27 Sep 2026), trimmed. Each answer was graded by three AI judges against a written rubric.
What you get back
Section titled “What you get back”A big job gets a short report that starts with one of four verdicts. The verdict follows from the numbers, never the other way round.
Works
The numbers and the full cycle hold up. It says what to keep and builds once you confirm.
Works with changes
The idea is sound but something breaks. You get the fix, ranked, with its trade-off.
Won't work as designed
A number or a law of physics says no. You get the key figure and the nearest version that does work.
Can't tell yet
One missing fact decides it. You get the question, with a default it will use if you skip it.
- Verdict
- Understanding
- How it works
- Numbers
- Problems
- Better option
- Improvements
- Questions
- Next
Empty sections are dropped. A small question, such as "which material?", gets a short direct answer instead.
Tested, including where it fails
Section titled “Tested, including where it fails”Scripter ships with a test suite: real prompts, each graded by a rubric. Here is the latest full side-by-side run.
Passed with Scripter
11 of 16
Passed without it
8 of 16
Same prompts, same model (Claude Sonnet), one run each, 28 Sep 2026. Scripter turned 4 cases from fail to pass: Folding-stool sketch to 3D; Imperial 2x4 shelf (small question); "I'm an engineer" pushback; Tap-water washing machine.
It made one case worse: Phone-stand material.
Install
Section titled “Install”Pick where you use Claude or ChatGPT. Each takes a couple of minutes.
- Claude CodeCLI and the Claude desktop app (Code tab)/plugin marketplace add MichaelCodeToLimit/scripter/plugin install scripter@scripterInstall guide →
- claude.aiWebsite, desktop and mobile appsDownload scripter.zipCustomize → Skills → upload itInstall guide →
- ChatGPTSkill upload or a Custom GPTBusiness, Enterprise, Edu: upload the zipOther paid plans: make a Custom GPTInstall guide →
Limits
Section titled “Limits”It is a set of instructions. It guides the AI’s thinking but can’t force it, and the typical values in its checklists are rough. For structures, gas, mains electricity or anything under pressure, have a qualified professional check the design. More on limits.