A hands-on experiment with Claude and a QWERTY-equipped TEC-1G retro single-board computer produced working Z80 assembly for a Wordle-style game—plus a few instructive mistakes along the way.
Could Claude write Z80 assembly that actually runs on retro hardware? One maker’s experiment suggests the answer is yes—at least when a skilled human steers the process.
The project centers on a toolchain built around [Ready Z80] and a TEC-1G single-board computer. a modern retrocomputer inspired by the TEC-1 design popularized in the 1980s.. The setup is intentionally humble: the base system is designed around a hex keypad. but the builder uses a QWERTY keyboard add-on. and he corrects the model when it makes assumptions.. That back-and-forth matters. because LLMs often project a “best guess” view of the target platform unless the operator actively tightens the constraints.
This is where Betteridge’s law creeps in: if a headline asks whether something can do something. the answer often ends up being “not really”—or at least “not reliably.” Yet the maker’s approach avoids treating Claude like an autonomous programmer.. Instead of dropping a single broad prompt and hoping for a finished program. [Ready Z80] walks Claude through the requirements in steps.. The result is a working game rather than a pile of plausible-but-broken code.
A key detail in the workflow is how tightly the human frames the problem.. LLMs can be fluent in code syntax. but retro assembly programming is unforgiving: a missing instruction. an incorrect calling convention. or a wrong memory layout can derail everything.. In the experiment, Claude initially produced some nonexistent instructions—overconfidence with legacy instruction sets is a familiar pattern.. The difference is that the builder catches the error. corrects Claude. and the model adjusts instead of insisting the mistake is real.
That “summer intern” feeling—handy, quick to draft, but needing direction—shows up repeatedly in the output.. The prompts are not vague.. The builder specifies what each part of the program must do, which reduces the model’s freedom to improvise.. It’s a reminder that for low-level work. correctness depends less on general intelligence and more on disciplined engineering: verifying assumptions. checking the instruction set. and testing on real hardware.
From a hardware perspective, the TEC-1G is a meaningful proving ground.. Emulation can hide problems; real systems force you to deal with input/output realities. timing. and the quirks of the specific platform.. The project’s focus on a Wordle-style game is also telling.. Wordle-like logic is structured. interactive. and stateful—ideal for stressing code paths that must handle repeated user input. display updates. and internal game state without the safety nets common in higher-level languages.
Why does this matter beyond retro hobby bragging rights?. Because the broader lesson applies to today’s software world: LLMs can accelerate drafting and scaffolding. but they don’t eliminate the need for technical literacy.. For engineers working close to the metal—embedded systems. firmware. performance-critical routines. or even security tooling—these assistants become most useful when paired with verification loops.. The tool may “feel” faster. but studies often find the workflow can also slow down if the operator spends time untangling confident errors.. In this case, the outcome is positive, but it doesn’t look like magic.. It looks like assisted development.
The experiment also hints at where the technology is headed.. As models improve, you’d expect fewer hallucinated instructions and better alignment with specific architectures.. But even then, assembly programming will remain a place where human review is non-negotiable.. The operator’s role—defining constraints, correcting misunderstandings, and validating on hardware—doesn’t just patch gaps.. It becomes the quality system.
For makers and developers watching this unfold. the takeaway is practical: if you want LLM help with legacy code or constrained systems. structure the conversation like a technical plan. not like a request for a finished miracle.. Give the model clear stepping stones, validate what it outputs, and assume it will occasionally wander.. When you do, you may get something rare for niche code—working software—without having to start from scratch.
Boost Portable Solar Panels Up to 30%: 11 Practical Tips
Cursor eyes $2B+ funding as enterprise AI coding accelerates
Dyson HushJet mini fan review: powerful—and louder than expected