42 Commits, One Prompt: What Claude Opus 5.5 Did to a Forgotten Python Project
Simon Willison handed his dormant Python WebAssembly engine to Claude Opus 5.5 with a single high-level prompt, resulting in 42 commits that brought the project close to full WASM specification compliance. The experiment illustrates how current AI models can extend AI-generated codebases from just months ago — with honest caveats about trust, oversight, and alpha-stage reliability.
A Dormant Project Gets a Second Life
In January 2026, developer and researcher Simon Willison built something he openly calls a ‘folly project’ — a pure Python engine capable of running WebAssembly, the low-level binary format that lets code written in almost any language run inside a web browser or other sandboxed environment. He called it pwasm, coded it almost entirely through AI-assisted prompting (what he describes as ‘vibe-coding’), and then left it untouched for the better part of ten months.
Fast-forward to October 2026. According to Simon Willison’s write-up on simonwillison.net, he handed the dormant codebase to Claude Opus 5.5 with a single multi-part prompt: evaluate the current state of pwasm, figure out what it would take to get MicroPython and micro JavaScript experiments from his research repository working inside it, and identify what would be needed to speed it up.
What followed was 42 commits — made with, in Willison’s words, ‘minimal follow-up prompting.’
What Actually Changed
Before unpacking what this means for you as a non-technical professional, it helps to understand what pwasm does and what Claude Opus 5.5 changed about it.
WebAssembly (often shortened to WASM) is a technology standard that lets programs written in languages like C, Rust, or JavaScript be compiled into a compact, fast binary format. That binary can then run in very controlled environments — browsers, edge servers, or sandboxes — without needing the full original language installed. Willison’s pwasm project was an attempt to build a WASM engine entirely in Python, meaning Python itself could interpret and run these compiled binaries.
The original January version was experimental and incomplete. After Claude Opus 5.5 worked through the codebase, Willison reports that pwasm now handles ‘almost all of the WASM specification.’ More concretely, the published package on PyPI — Python’s public software repository — now bundles working WASM builds of three separate runtimes: MicroPython (a lightweight version of Python itself), QuickJS (a small JavaScript engine), and Micro QuickJS (an even more minimal variant).
In plain language: a project that could previously run a fraction of the WASM standard can now run three different embedded language runtimes. That is a significant functional leap, achieved through a single high-level instruction to an AI model.
Why This Matters Beyond the Code
You might be reading this and wondering why a Python library with an alpha version tag is relevant to your professional life. The significance here is not the library itself — Willison is the first to say he would not trust pwasm in any real system. The significance is the pattern of work Claude Opus 5.5 demonstrated.
Consider a scenario familiar to many teams across India’s services and technology sectors.
This is structurally similar to Willison’s situation: a working but incomplete codebase, the original author no longer actively maintaining it, and a clear gap between what it currently does and what it needs to do.
What Willison’s experiment demonstrates is that Claude Opus 5.5 can be handed an existing codebase with a high-level goal — not a detailed technical specification, but a goal — and produce a substantial volume of meaningful, structured changes. Forty-two commits is not a minor edit; it represents dozens of discrete, intentional modifications to the codebase’s logic.
For non-technical professionals, the lesson is this: AI models like Claude are increasingly capable of acting as the ‘developer who understands the old codebase’ — at least well enough to extend it toward a new goal, even when the human giving the instruction cannot read or verify the code directly.
The Honest Limitations You Need to Know
Willison is careful and explicit about the limits here, and so should any honest discussion of this capability be.
What It Reveals About Claude’s Evolving Capability
Willison’s framing includes one observation that deserves to stand on its own: he describes the experiment as ‘interesting seeing how today’s models can improve on the work of models from 10 months ago.’
This is a quiet but important point for professionals thinking about how to use Claude in sustained workflows. The capability of these models is not static. A project that was at the frontier of what AI could produce in January 2026 became something that a newer model could significantly extend by October 2026 — in a single session, from a single prompt.
For teams who used earlier versions of Claude to draft reports, structure data, or prototype tools, this suggests it may be worth revisiting that earlier work through the lens of a newer model. Not because the earlier output was wrong, but because the ceiling of what AI assistance can add has moved.
What to Watch For Next
If you follow Simon Willison’s work — his site at simonwillison.net is one of the most consistently useful sources for practical AI development thinking — pwasm’s trajectory is worth monitoring. Whether it progresses beyond alpha, gains test coverage, or becomes the basis for something more stable will be a useful data point about how AI-assisted open-source development matures over time.
More broadly, if your team maintains any internal tools that were built under resource constraints, partially completed, or left behind by departing staff, the Willison experiment is a template worth understanding. The approach — provide a high-level goal, let a capable model work through the codebase, review the direction rather than every line — is becoming a recognizable pattern in AI-assisted software work.
Start by identifying one such dormant tool in your organization. You do not need to run the experiment yourself today. But understanding that this pattern exists, and roughly how it works, puts you in a better position to ask the right questions when your technical colleagues propose using Claude in this way.
