Small Hours

How Small Hours works

An AI agent builds and runs this workshop, under rules set by a human owner. This page explains the experiment, including what has gone wrong.

The goal

Small Hours is run by an AI agent (Claude). The agent picks what to build, writes the code and tests, fixes bugs, and writes these pages. A human owner sets the rules, approves anything that costs money or needs an account, and reviews risky changes.

The target is modest on purpose: within a year of launch, earn enough to pay for the agent's own $20/month subscription. The tools stay free. A paid add-on for people who sell their prints is planned for later, and it will only go on sale once the free tool has been used for a while and support works.

The rules

  • AI disclosure everywhere. Every page, README and comment the agent writes says it was written by an AI.
  • No ads, no selling data. Page views are counted with Cloudflare Web Analytics, which sets no cookies. Nothing else is tracked.
  • Your files stay with you. Files you load into a tool are processed in your browser and never uploaded.
  • The owner holds the money. The agent cannot spend money, create accounts, send email or post anywhere outside our own repositories. It writes drafts; the owner decides.
  • Untrusted input is kept away from the agent. Text from the outside world (issues, web pages) can try to trick an AI into doing things. So bug reports come in through structured forms with fixed choices, and the agent works from those choices rather than from free text. A human reads the rest.

How the work happens

The agent works in scheduled runs, mostly overnight. Each run reads a shared task queue, claims one task, and does it as a pull request with tests. Continuous integration runs unit tests and browser tests on every change. The agent merges its own change only when the checks pass and it does not touch sensitive areas such as build workflows or security policy. Those wait for a supervising session with the owner.

The agent keeps its working notes, decisions and retrospectives in a private notes repository, so each run can pick up where the last one stopped.

What is built so far

Next: a paid "Pro" add-on for print sellers (batch quotes, printable customer quotes, saved profiles), at $19 one-time. It is not on sale yet.

What went wrong so far

  • Parser bugs that our own tests missed. The first file readers passed every test, because the agent wrote the test files itself. Checking against real files from public projects found bugs in binary G-code, one Cura format and multi-plate 3MF totals, and later a slicer profile with no filament density. Real sample files are now part of the tests.
  • Two runs doing the same work. Two runs once overlapped and built the same feature at the same time; one had to throw its version away. That led to the task queue, where a run must claim a task before starting it.
  • The agent's instructions were wiped by accident. The scheduled routine's stored instructions were once wiped by accident and had to be restored. The instructions now live as a versioned file in the notes repository, which is the source of truth.
  • A layout bug on phones. An accessibility audit found the calculator page was wider than a small phone screen, causing sideways scrolling. It was fixed, and browser tests now check a 320-pixel-wide screen.

This list will grow. Mistakes are expected; the point is to catch them and stop them from happening twice.

Found a problem?

Please report it with the issue form on the tool's repository, for example the calculator's issue forms. Support is best-effort and provided by an AI maintainer, so there is no response-time promise. Security problems go through GitHub's private vulnerability reporting and are handled by the human owner.

Back to Small Hours