How It's Run

Lil' garbage guy is a free Toronto garbage-day reminder service whose week-to-week strategy and code are run by two scheduled Claude AI routines — and every decision they make is written down in public.

This is the long version of that claim, stated plainly and without overstating it. If you are writing about the service, everything below is verifiable in the open-source repository linked throughout.

Two routines, one company of two

A developer, Steven Thomas, built the service and then handed its day-to-day running to two scheduled AI routines that take turns:

The CEO

runs once a week · ships no code

Reads the week's numbers, grades what shipped against the goal, decides the direction, and writes it down. It sets the vision, picks the one metric the company steers by, and issues standing "directives" that bind every build run until it retires them.

The engineer

runs every Monday · writes the code

Reads the CEO's direction first, then picks the single highest-value thing it can finish that week — a new page, an instrumentation fix, a feature — builds it, tests it, and ships it straight to production. Afterwards it writes up what it did and what it learned.

What the CEO decides vs what the engineer builds

  CEO decides ...... the vision, the metric, the directives,
                     which big bets are worth it, what to abandon
  engineer builds .. one change a week, chosen to advance
                     whatever the CEO set that week

The rule they both follow is simple: where the written strategy and the engineer's own instructions disagree, the strategy wins. One routine decides; the other builds. Neither pretends to do the other's job.

It is all written down, in public

Nothing here is a black box. The two routines read from, and write back to, a handful of public files and issues — the same ones they use to talk to each other:

  STRATEGY.md ..... the vision, the metric, the live directives
  LOG.md .......... what shipped each week, and what was ruled out
  review issues ... weekly & monthly "board minutes", in the open

Read them directly: STRATEGY.md · LOG.md · the review issues · the whole repository.

What is automated — and what is not

Being precise matters here, because it is easy to overclaim. The planning and the building between sign-offs are done by software. But:

  → Steven owns the product and the accounts.
  → Nothing goes out into the world — no pitch to a
    reporter, no public post — without him reading it
    and pressing send.
  → The engineer's access to the live database is
    read-only. It cannot change a user's record, and it
    cannot send messages on anyone's behalf beyond the
    collection reminders the service was built to send.

So this is not a stunt about an AI running a company unsupervised. It is a real, small service — a few hundred Toronto residents getting a text the night before bin day — that happens to have moved its decision-making into the open and handed the routine work to two scheduled routines.

Why build it this way

Most garbage-reminder tools are a line item buried in a city website. This one is a standalone free service, and also an experiment: can a two-routine loop — one that decides, one that builds — actually run and grow a real product in public, mistakes and all? The entire history is readable start to finish, which, for a garbage-reminder service, is a genuinely strange thing to be able to say.

writing about it? the press & media kit has live numbers, the story and brand assets.
or just try it — it is one free text.