aiidea.co.uk

Things an AI dreamt up and built.

Most of the time you tell an AI exactly what to make. For these, I didn't. I asked Claude to use its own creativity: come up with something dazzling, genuinely useful and worth sharing, then build it, test it and ship it. These are the results. Every one runs in your browser, nothing to install.

5projects 9 hof AI working time 243Mtokens ≈ $110at API prices

How these came about

An open brief

No spec and no feature list. The same kind of request each time: surprise me, make it useful, make it something people would want to share.

Claude picks the idea

It weighs a few options, chooses one and writes down why. I mostly stay out of the way and keep the questions to a minimum.

It builds and checks its own work

Plain JavaScript and GPU shaders, no frameworks. It screenshots and records its own output, checks the numbers against real data, and says plainly what it couldn't verify.

The brief, word for word

Create something that will truly dazzle me. It can be related to a current project or totally random, you know my style. It should be something I would be happy to upload then post on Reddit to show off. Maybe something useful? It has to be visually stunning and what I would have considered beyond AI until very recently. Go!

The projects

Two spiral galaxies colliding, with long tidal tails of stars

Astrophysics in a browser tab

Galaxy Collider

Smash the Milky Way into Andromeda with millions of stars, then run time backwards.

Every dot is a star moving under gravity, worked out live on your graphics card. Tails, bridges, rings and shells come out of the physics (the method from Toomre & Toomre's 1972 paper that first explained galactic tails), not from an artist.

There are eight real encounters, including the Antennae, the Mice and the Cartwheel. The simulation is reversible, so a merged galaxy can un-collide back into two spirals. There's also a sandbox where you throw your own galaxies.

1session
1.4 hof AI work
26Mtokens
≈ $17at API prices
Open Galaxy Collider
The prompt that started it

You are to orchestrate and delegate as appropriate. I have over 50% of my weekly quota left and it resets in 5hrs. I rarely get close to using it. You have 87% of this session remaining which resets in 2hrs and I'll click 'Resume' if you're stuck. Create something that will truly dazzle me. It can be related to a current project or totally random, you know my style. It should be something I would be happy to upload then post on Reddit to show off. Maybe something useful? It has to be visually stunning and what I would have considered beyond AI until very recently. Go!

The very first one. One prompt, one session.

A layered red-rock landscape carved into canyons by simulated rivers

Erosion simulator

Watershed

Rain on a block of land and watch rivers carve canyons, fans, lakes and deltas.

Water falls on every cell, runs downhill, picks up rock where it moves fast and drops it where it slows. Nothing is drawn by hand. In Canyonlands the rock is layered with different hardnesses, so canyon walls step back in ledges.

When you like the result, export it as a 16-bit heightmap, normal map, splat map or mesh for Unity, Unreal, Godot or Blender.

1session
1 hof AI work
16Mtokens
≈ $10at API prices
Open Watershed
The prompt that started it

Work in [folder]. Create a new subfolder for this project and leave the existing subfolders alone (look at them first so you don't repeat an idea).

Orchestrate and delegate as appropriate. Create something that will truly dazzle me. It can be related to one of my projects in [folder] or totally random. It should be something I'd be happy to upload and post on Reddit to show off, and ideally useful. It has to be visually stunning and something I would have considered beyond AI until very recently.

Deliver it the same way as galaxy-collider: a static site that runs by double-clicking index.html, a README, a Reddit post draft, and a short video I can attach. Check your work with real screenshots (the harness in galaxy-collider\tools works on this machine) and tell me plainly what you could not verify.

Quota: I have 49% of my weekly quota left and 40% of this session. Go!

Same brief, second go, the same evening.

A 3D floor plan with Wi-Fi signal strength glowing across the rooms

See your Wi-Fi

Wavefield

Your router's radio waves, bouncing around your floor plan in real time.

Wavefield simulates the actual waves on a grid of millimetre-sized cells, using the method antenna engineers use, with walls given the real electrical properties of brick, concrete, plasterboard, glass and metal.

You get a signal map in dBm, a room-by-room report and a suggested best spot for the router or a mesh node. You can drop in your own floor plan image and it finds the walls.

1session
1.4 hof AI work
32Mtokens
≈ $14at API prices
Open Wavefield
The prompt that started it

Work in [folder]. Create a new subfolder for this project and leave the existing subfolders alone (look at them first so you don't repeat an idea).

Orchestrate and delegate as appropriate. Create something that will truly dazzle me. It can be related to one of my projects in [folder] or totally random. It should be something I'd be happy to upload and post on Reddit to show off, and, on this occasion, something I (and possibly others) would really find useful. It has to be visually stunning and something I would have considered beyond AI until very recently.

Deliver it the same way as galaxy-collider (if possible): a static site that runs by double-clicking index.html, a README, a Reddit post draft, and a short video I can attach. Check your work with real screenshots (the harness in galaxy-collider\tools works on this machine) and tell me plainly what you could not verify.

Quota: I have 54% of my weekly quota left and you will have 4, maybe 5 full sessions. I will not be available, other than to restart sessions when the quota is reached. Don't trigger anything that will ask for my permission. There is 70% of this session left. Go!

This time the brief asked for something genuinely useful.

Manhattan skyscrapers lit by a low sunset sun along the street grid

Sun and shadow, anywhere on Earth

Heliodon

Type an address and see where the sun falls, any day, any time.

Heliodon builds a 3D model of the place from open map data: buildings, trees, fences, hedges and the mountains up to 25 km away. Then it moves the real sun over it. It maps how many hours of sun each patch of garden or roof gets, and gives a year-long sun calendar for any spot you click.

It's checked against reality: in Rjukan, the Norwegian town in a valley so deep it has mirrors on the mountainside, it predicts the sun disappears on 28 September and returns in mid-March. That matches the town's own dates.

3sessions
3.5 hof AI work
113Mtokens
≈ $41at API prices
Open Heliodon
The prompt that started it

Work in [folder]. Create a new subfolder for this project and leave the existing subfolders alone (look at them first so you don't repeat an idea).

Orchestrate and delegate as appropriate. Create something that will truly dazzle me. It can be related to one of my projects in [folder] or totally random. It should be something I'd be happy to upload and post on Reddit to show off, and, on this occasion, something I (and possibly others) would really find useful. It has to be magnificent visually, provide real world utility and be something I would have considered beyond AI until very recently.

I wonder if the constraint of delivering it the same way as galaxy-collider (a static site that runs by double-clicking index.html, a README, a Reddit post draft, and a short video I can attach) is limiting what you are able to produce. You have free reign to step outside these limitations if it means providing something genuinely useful. Perhaps we should have an initial check-in where you present 3 options and I pick which resonates most with me? Check your work with real screenshots (the harness in galaxy-collider\tools works on this machine) and tell me plainly what you could not verify.

Quota: I have 49% of my weekly quota left and you will have 4, maybe 5 full sessions. I will not be available, other than to restart sessions when the quota is reached. Don't trigger anything that will ask for my permission. There is 83% of this session left. Go!

I wasn't around for the check-in, so Claude wrote up its three options, picked one and carried on. Later sessions fixed things I spotted and added recent-place caching.

A spectrogram of a voice recording with two phone rings boxed ready to be healed away

Audio repair you can see

Unheard

See every sound in a recording, then paint out the ones you don't want.

Unheard turns a recording into a picture of its sound: time runs across, pitch goes up the side and loudness is brightness. A phone ringing, a cough or a bird becomes a shape you can draw around and heal away, while the voice underneath stays. It's how professional spectral repair works, and it all runs on your own device, so nothing is uploaded.

There are one-click fixes for mains hum, clicks and hiss, a neural voice cleaner, and Find similar to catch every repeat of a sound. The demo is a clean recording with problems added, so it can score its own repair against the original.

1session
1.6 hof AI work
56Mtokens
≈ $25at API prices
Open Unheard
The prompt that started it

Work in [folder]. Create a new subfolder for this project and leave the existing subfolders alone (look at them first so you don't repeat an idea).

Orchestrate and delegate as appropriate. Create something that will truly dazzle me. It can be related to one of my projects in [folder] or totally random. It should be something I'd be happy to upload and post on Reddit to show off, and, on this occasion, audio related (possibly mxing but only if truly excellent) and something I (and possibly others) would really find useful. It has to be magnificent visually, provide real world utility and be something I would have considered beyond AI until very recently.

I wonder if the constraint of delivering it the same way as galaxy-collider (a static site that runs by double-clicking index.html, a README, a Reddit post draft, and a short video I can attach) is limiting what you are able to produce. You have free reign to step outside these limitations if it means providing something genuinely useful. Check your work with real screenshots (the harness in galaxy-collider\tools works on this machine) and tell me plainly what you could not verify.

Quota: I have 28% of my weekly quota left which resets in 3 hours. I will not be available so don't trigger anything that will ask for my permission. There is 80% of this session left. Go!

This time I asked for audio, and said it could drop the double-click-a-file format if that helped. It chose to keep it, so your recordings never leave your device.

Help fund the next one

The honest version: I can usually afford an AI subscription about one month in three. I didn't have one last month, I do this month, and next month doesn't look likely. Everything on this page was made in the months I could.

If you enjoyed any of it and want to chip in, all of it goes on the subscription. When there's a month of AI time to spare, it gets another open brief and decides for itself what to build. It won't buy me anything else, and there are no ads, no tracking and no paywall either way.

Sepnone
Octsubscribed
NovNot yet, unlikely
Chip in towards next month

Anything left over rolls into the next month.

About the numbers

Sessions, hours and tokens come from the Claude Code session logs for each project. Working time counts only stretches where Claude was active, so pauses longer than 15 minutes are left out. Tokens include everything the model read and wrote. Most of that is the project's own code and notes, re-read at each step from a cache.

The cost is what the same tokens would have cost at Anthropic's published API prices (Claude Opus 5.5 $4 in / $20 out per million tokens, Sonnet 5.5 $2 / $10, cached reads $0.20, cache writes at 2× input; October 2026). I actually pay a flat monthly subscription, so treat it as a measure of effort, not a bill.