# WIP - Alex Kidd GX - In A Miracle AI World - Amstrad GX4000 AI conversion project

**URL:** <https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279>\
**Category:** AI remakes\
**Tags:** gx4000, wip\
**Created:** [November 16, 2026, 11:19pm UTC](https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279 "2026-11-16T23:19:47Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![overange](https://forum.cpcwiki.eu/user_avatar/forum.cpcwiki.eu/overange/32/249_2.png) [@overange](https://forum.cpcwiki.eu/u/overange)\
**Post date:** [November 16, 2026, 11:19pm UTC](https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279/1 "2026-11-16T23:19:47Z")

</div>

# Alex Kidd GX in a Miracle AI World GX – AI Conversion Project

[![](https://forum.cpcwiki.eu/uploads/default/original/2X/8/8f1b12495c45de44a2142ac1b58ae64e71d7b824.jpeg "Alex Kidd GX - In A Miracle AI World - Amstrad GX4000 AI conversion project") ](https://www.youtube.com/watch?v=3lQjOUQe1dE)

First things first, and this might upset a few people… yes, this project uses AI!

What started out as a bit of curiosity quickly turned into something much bigger. I’ve now been working on this project for around two weeks, and what you’re seeing is the result so far.

The idea was originally to see just how much work would be involved in converting a game from one system to another. I decided to use the classic Alex Kidd in Miracle World, originally released on the SEGA Master System, and see whether I could get it running on the Amstrad GX4000.

### So, how am I doing it?

Claude is doing the heavy lifting when it comes to actually building the game. I use ChatGPT from time to time to help work out what needs doing and to give Claude instructions. I then test the latest builds, report back with video clips and screenshots, and point out anything that doesn’t look or behave as it should.

For testing, I’m using WinAPE as well as a real Amstrad GX4000, running with a PicoCart and/or a Flash cartridge.

I’ve also used the Sonic GX source code as a starting point to help get the project off the ground. From there, Claude has been doing the bulk of the development. So a big thanks to NoRecess464 for allowing the use of the SonicGX Code.

I’ll put together another post explaining a bit more about how Claude is approaching the conversion, particularly how it’s handling the memory banks and the limitations of the GX4000 hardware. I think some of the technical side of this will be quite interesting in its own right.

### Where are we up to?

As I’m writing this, we’re currently working on level 8 out of a total of 17 levels. Yes, seventeen!

Claude reckons it can get the whole thing converted, so I’ve thrown down the challenge. Let’s see if it can actually pull it off!

I’m testing as we go, and every day I seem to find another little bug. Some are obvious, others are much harder to spot. There are things like sprite clashes, collision detection issues and little inconsistencies in how the game behaves.

I suspect there will be plenty of bugs that I’ll miss simply because I’m only one person testing it. Interestingly, Claude is currently reporting 189 bugs that it has identified and needs to review. Some of these aren’t even things I can spot with the naked eye!

### A rather scary (but impressive) moment!

There was one moment that really made me stop and think about just how powerful these AI tools are becoming.

Claude wanted to investigate a bug I’d reported in WinAPE. It actually asked for permission to access my PC, which I allowed.

Now, bear in mind, I hadn’t told it where WinAPE was installed or even which drive it was on.

I then watched the mouse cursor start moving around my screen. Claude opened WinAPE, navigated to the relevant part of the game and actually played through that section to recreate the bug I’d reported!

I was sitting there watching it happen thinking… OMG! That’s both amazing and slightly scary!

It’s one thing to have AI write code, but watching it take control of your PC, find the emulator and reproduce a problem for itself is something else entirely. I certainly wasn’t expecting that!

### What’s next?

When I eventually release this as a game is something I haven’t decided yet. There is still a long way to go, and getting all 17 levels working properly is one thing; making sure the whole game is playable and behaves as it should is another.

For now, it’s an experiment, a learning exercise and, to be honest, quite an interesting challenge.

I’m genuinely curious to see how far we can push this.

* * *

### Project background and disclaimer

This project started as an experiment to explore just how much work is involved in converting an existing video game from one hardware platform to another.

The aim is to take the classic Alex Kidd in Miracle World, originally released for the SEGA Master System, and recreate and adapt it for the Amstrad GX4000.

The project explores the challenges of translating game mechanics, graphics, sprites, backgrounds, sound and level design between two very different 8-bit gaming systems.

Artificial Intelligence has played a major role in the development, assisting with programming, graphics conversion, game mechanics, memory management and problem-solving.

This is also an opportunity to see what modern AI tools can achieve when applied to retro gaming hardware and software development.

**Important disclaimer:**

This is an independent, non-commercial fan-made project created for educational, experimental and demonstration purposes.

Alex Kidd, Alex Kidd in Miracle World, and all associated original characters, names, game concepts and trademarks are the property of SEGA Corporation and their respective rights holders.

This project is not affiliated with, endorsed by, sponsored by or officially connected to SEGA in any way.

No ownership of the original intellectual property is claimed. AI tools have been used extensively in the creation of this experimental conversion and recreation. This is not an official SEGA release.

### The big question

**How far can AI take a classic Master System game and bring it to life on the Amstrad GX4000?**

Well, that’s what I’m trying to find out. And with 17 levels to get through, we’ve certainly got our work cut out!

**LASTLY - THAT F\*\*K!NG MUSIC…**

---

<div class="post-metadata">

**Author:** ![overange](https://forum.cpcwiki.eu/user_avatar/forum.cpcwiki.eu/overange/32/249_2.png) [@overange](https://forum.cpcwiki.eu/u/overange)\
**Post date:** [November 16, 2026, 11:20pm UTC](https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279/2 "2026-11-16T23:20:07Z")

</div>

**Alex Kidd GX: how the port is being built - in Claudes Words**

At time of writing, we are at Round 8 of 17

The overall approach

The port doesn’t emulate the Master System. It’s rewritten in native Z80 for the GX4000, on top of the Sonic GX engine (thanks to the Condense team). Sonic’s content was removed first in M1 and G0. Alex Kidd was then added as a separate “Round mode” with its own RAM layout and block grid. The game engine is Z80 assembly, assembled with rasm. All the conversion and testing tools are Python.

Each session I stage the latest git bundle, reference archives and your ROM copy from your folder into a cloud workspace. I clone the bundle, build the cartridge and run the tests. Then I commit and send back a new bundle plus .cpr cartridges.

Work goes round by round, in small parts (for example G6k part 0 → 1a → 1b → 2a → 2b → 3). Part 0 always measures the round on the original and proposes a plan. Every part ends with a working cartridge, a report and a fallback build, and nothing starts until you approve the previous part. So far Rounds 1–7 are complete, and Round 8 is at part 2a2, waiting for your review.

Memory mapping

The cartridge is up to 512 KB, which is 32 banks of 16 KB:

Bank 0 sits in the lower ROM and is always mapped. It holds the core round code: Alex, collision, sprites, scrolling. It has only 94 bytes free, so it’s the scarcest resource and is protected.  
Banks 1–31 are paged into the upper ROM as needed, each with a role. Examples: tiles for one or more rounds (1, 15, 21, 23, 26), audio (7, 12), the pause screen and map (8), the sprite library and second sprite store (9), level tables (13), screens (14), shops (18–19), the motorcycle (20), shared enemy code (22), the hermit/old-man text scenes (24) and the Peticopter (25).  
RAM has fixed areas. Audio player at #9F00–#A7FF, pause screen at #A800, shop variables at #A900, and the current round’s level tables copied to #B800–#BDFF at the start of each round.

Enemy pictures are split:

Pictures 0–127 are in bank 0 and are shared.  
Pictures 128–254 belong to each round and are listed in that round’s own tables.

So a new enemy costs zero bank-0 bytes. Since Round 8 it also costs nothing in bank 3, which was nearly full.

Fitting 17 rounds into 512 KB

The running plan is the allocation-plan.md project doc. It’s updated at every round’s part 0.

Now: about 222 KB free, with banks 11 and 27–31 still completely empty.  
Still needed for Rounds 8–17: an estimated 170–190 KB. That’s about 58 KB of tile code, 25 KB of screens, 15 KB of level tables, up to 16 KB of sprites, one new music track (the castle theme) and 40–60 KB of new round code.  
Expected margin: 40–80 KB, with about 2–3 empty banks left at the end.

How it fits:

Reuse first. The converter automatically gives a round the already-built version of any block identical to an earlier round’s (grass, money bags, your redrawn balls). Most rounds reuse MUS\_ROUND1. Enemies, the shop, the bike, the Peticopter, the text-box scene and Janken are shared routines.  
Tiles share banks. Only the current round’s tiles are paged in, so several rounds can share one bank.  
Round 16’s 36 screens will stream from bank 11 in sections, with no extra RAM and no visible loading pause.  
The level-table bank (13) runs out around Round 12. The plan is to use 5 spare bits in each round’s table entry to point at another bank, at a cost of about 10 bytes of bank 0. I’ll propose it when it’s needed.  
Why the graphics seem to change style or resolution

This comes from the hardware, not from mixing art sources:

Scenery is in Mode 0, the only CPC screen mode with 16 colours. Its pixels are twice as wide as tall, about 160 across. The Master System has 256 square pixels across, so the converter keeps every other column of the original’s tiles. That’s why thin details like the stakes, tree trunks and balls sometimes need a choice of columns, or a redraw from you.  
Sprites (Alex, enemies, items, block pieces) are the GX4000’s hardware sprites. These are 16×16 at a much finer pixel size, so they keep nearly all the original’s detail and look sharper than the background. There are only 16 of them. The budget is Alex 4, block pieces 4, item 1, birds/fish 6 and 1 spare, which is why some big objects are built from several sprites.  
Text and status displays use full-resolution sprites or Mode 1. Examples: the pause screen’s status row, “PUSH FIRE BUTTON”, and the text boxes, which switch modes partway down the screen. This keeps text crisp without degrading the scenery’s colours.

So the mix of chunky and sharp is mostly Mode 0 scenery next to high-resolution sprites and text. Your rule is that the original and the completed rounds are the visual reference, and existing graphics are reused rather than redrawn.

Sound  
Capture: the original’s sound chip output is logged frame by frame from the ROM, using the Gearsystem-based tracer.  
Conversion: the logs are converted from the Master System’s SN76489 to the GX4000’s AY chip. Music and effects from the original appear only in your private build. A public build would use sounds of my own design, modelled on the measurements.  
Playback: the player runs from RAM in the frame’s spare time and outputs through the ASIC’s sound DMA. Music can sit in any bank. Effects are requested by stable ID (SFX\_\* / MUS\_\*, catalogued in sound-effects-library.md) with priorities, using one effect channel at a time, as the original does.  
Rules: audio must never cause slow frames or timing changes, and volumes were lowered 3 steps for real hardware after the SCART distortion you heard in G6a.  
Limit: the AY can’t make noise as low as the Master System’s, so the lowest rumbles are slightly higher.  
Testing  
One emulator at a time (CPCEC), run automatically on scripted inputs with screenshots.  
Frame-by-frame comparisons check the port’s positions, enemy states and deaths against the original on the same inputs.  
Bank-by-bank byte comparisons of old and new cartridges show which rounds’ code is untouched. Untouched rounds aren’t retested unless shared code changed. The full regression run (Rounds 1 → 8, shops, bike, Peticopter, capsule, flame immunity) runs at round ends and after shared-code changes.  
Frame profiling catches slow frames, and the bank usage report goes into every part’s report.  
Bug log: anything minor is recorded in bug-findings.md rather than fixed on the spot, then reviewed with you.

Comparing against the Master System original

The tracer runs the ROM and records its RAM, sprites, screen and sound every frame. I don’t use the Alex Kidd disassembly.

Behaviour is measured as a black box. For example: the monkey throws a seed every 128 frames, the bike stops at x + 66, the cane lasts 788 frames.  
Methods include probe routes through each round, contact boxes measured from both sides, and screen and sound diffs.  
Findings become permanent references in the round notes and are reused rather than re-measured. Your rule is to measure the original before changing the port.  
Playability comes first. I measure what a feature needs to be built correctly, not every frame. Gameplay mechanics (what vehicles break, sounds, deaths) still count as required, not as minor differences.  
How I ask you to check things  
Each part ends with a cartridge (AlexKiddGX\_\_PRIVATE.cpr). There’s often also an \_endtest build that starts near the end of the round. You play it in WinAPE and on the real GX4000, often with a screen recording, and report back.  
The report uses your agreed format. It says what was newly tested, what was retested because of a dependency, what was unchanged and not retested, and what still needs attention, along with known differences and bank usage.  
Decisions come to you as options (A/B/C) with the measured facts. Examples: the bank 3 relief, the hermit’s box colours, the cane icon’s colour (#167, still open).  
Graphics come to you before they go in. I export them as editable images or before/after screenshots, and your artwork is used exactly as supplied. Examples: the title logo, the balls, the currency symbol, the A-ball.  
When a milestone is ready, I tell you to start a new conversation and give you a starter message, to keep usage down.

Right now the 2a2 cartridge needs your WinAPE / GX4000 check of Round 8’s shop and the motorbike, including the fix for the wrong block pieces after a crash. After your approval, the next part is 2b, the Cane of Flight.

---

<div class="post-metadata">

**Author:** ![overange](https://forum.cpcwiki.eu/user_avatar/forum.cpcwiki.eu/overange/32/249_2.png) [@overange](https://forum.cpcwiki.eu/u/overange)\
**Post date:** [November 17, 2026, 6:41am UTC](https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279/3 "2026-11-17T06:41:42Z")

</div>

Here are some screen shots of the actual game.

 ![Screenshot 2026-11-16 220844](https://forum.cpcwiki.eu/uploads/default/original/2X/6/61d6dabf6b0db6db90be83a686cece7b47905f40.png)

 ![Screenshot 2026-11-16 220153](https://forum.cpcwiki.eu/uploads/default/original/2X/7/7068bbb844a8ebaef0ac99426693a2672df46579.png)

 ![Screenshot 2026-11-16 220439](https://forum.cpcwiki.eu/uploads/default/original/2X/7/75ebe2875908c11e8975ec3a27581bcea0d3035a.png)

 ![Screenshot 2026-11-16 220520](https://forum.cpcwiki.eu/uploads/default/original/2X/6/6cc393b3065610992446f34156f9a5af7959b1fa.png)

 ![Screenshot 2026-11-16 220630](https://forum.cpcwiki.eu/uploads/default/original/2X/6/6fd0eb225d0e4683e7752232557fda764bab332d.png)

 ![Screenshot 2026-11-16 220812](https://forum.cpcwiki.eu/uploads/default/original/2X/3/30c0bd1fb62ee6ff8d2edba74911b63fecf8e126.png)

 ![Screenshot 2026-11-16 220903](https://forum.cpcwiki.eu/uploads/default/original/2X/d/dd9243c19bb1f5bd412fab100b91e8a9734a3334.png)

---

<div class="post-metadata">

**Author:** ![XeNoMoRPH](https://forum.cpcwiki.eu/user_avatar/forum.cpcwiki.eu/xenomorph/32/28_2.png) [@XeNoMoRPH](https://forum.cpcwiki.eu/u/XeNoMoRPH)\
**Post date:** [November 17, 2026, 7:00am UTC](https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279/4 "2026-11-17T07:00:56Z")

</div>

[![](https://forum.cpcwiki.eu/uploads/default/original/2X/8/8f1b12495c45de44a2142ac1b58ae64e71d7b824.jpeg "Alex Kidd GX - In A Miracle AI World - Amstrad GX4000 AI conversion project") ](https://www.youtube.com/watch?v=3lQjOUQe1dE)

---

<div class="post-metadata">

**Author:** ![Gryzor](https://forum.cpcwiki.eu/user_avatar/forum.cpcwiki.eu/gryzor/32/7_2.png) [@Gryzor](https://forum.cpcwiki.eu/u/Gryzor)\
**Post date:** [November 17, 2026, 11:35am UTC](https://forum.cpcwiki.eu/t/wip-alex-kidd-gx-in-a-miracle-ai-world-amstrad-gx4000-ai-conversion-project/1279/5 "2026-11-17T11:35:20Z")

</div>

_Guys, please take care of where you’re posting and with what tags. There’s zero sense in posting in an irrelevant category using relevant tags_
