smol developer

There is no unkind way to put this and no reason to soften it. The project stopped receiving commits in 2024, and a code generator frozen against a moving field degrades without anyone touching it, because the assumptions baked into its prompts stopped matching how models behave. Whatever you got from it two years ago you will not get today.

It is worth an hour of your reading time as history, since a great deal of what you use now started here. Pick Aider for scaffolding and editing that is actually maintained. Do not adopt this.

Abandonment here does not announce itself with an error. The prompting strategy was tuned for the models available in 2023, and those prompts still run and still return something, so what you get is silently worse output rather than a failure you can act on. That is the more dangerous kind of stale, because nothing tells you to stop.

There is no version to pin your way out of. Done right, and worth preserving: a human sits between iterations by design, so nothing accumulates unreviewed across a whole generation run.

1. The input is prose, and the first output is not code but a plan naming the files that will exist, which makes the structural decision inspectable before any generation cost is incurred. 2. Each file is then produced with that plan as shared context, so consistency between files derives from a common artefact rather than from a growing conversation.

2. That separation of structure from content is now standard, which is the strongest evidence it was correct. No benchmark was published and none was claimed. The design predates tool use entirely, which is why it can only write and never check.

This is the purest example on the board of an idea that generated enormous value and captured none of it. Twelve thousand stars, wide influence, and no company, no hosted product and no revenue line was ever built on top. Every well-funded tool in this category shipped the concept within two quarters and the originator kept the citation.

Moat: none, and none was attempted, which is a legitimate choice rather than a failure. Likely path already happened: influence without capture, then dormancy. Position: none available. The instructive part is how quickly the idea escaped the repository.

An unmaintained dependency with no organisation behind it fails my procurement checklist at the first line, which is who answers when it breaks. Nobody does. There is no support agreement to sign, no security contact to notify, and no release cadence to plan around, so any team standardising on it has quietly taken on maintenance of a tool they did not write.

It produces new projects rather than changing existing ones, which is not the work sixty engineers spend their weeks on anyway. Onboarding cost is low and entirely wasted. Not yet, and there is no version of later.

MIT, and the thing worth taking is that it shipped as a library first. Importing smol_dev into my own script rather than shelling out to somebody's interface is how I want every tool in this category built, and it also exposed a standard agent interface at a time when almost nothing did. That was correct and mostly forgotten.

What kills it for me now is the model layer: two hosted vendors, no substitution, no local endpoint. So the honest use is lifting the planning function into something current. Fork is the wrong word; salvage is closer.