Standardizing eleven repos in a day (with an AI chief of staff)

I run two small companies and about a dozen codebases. Like most solo-ish technical founders, my repos had drifted: three different conventions for agent instructions, four competing code review setups, secrets committed in plaintext to a private repo, a live community site whose source existed in four repositories, and 143 stale branches on my flagship product. None of it was broken, exactly. All of it was friction.

Yesterday I handed the whole mess to an AI agent with a two-phase brief: audit everything and propose a plan — then stop. Only after I reviewed the findings did phase two start. By the end of the day, every repo followed one written standard, and I'd approved every irreversible step from my phone.

What actually got done

• One standard, eleven repos. Every repo now has a canonical AGENTS.md (instructions any AI tool can read), a CONTEXT.md facts file (pricing, integrations, do-not-touch lists — so agents load facts instead of hallucinating them), a real README, and CI.
• A four-repo merge. My community site's theme, ops docs, forum config, and design system became one monorepo — 413 commits of history preserved — and the trick was renaming the repo that already held the deploy secrets instead of recreating them.
• Secrets encrypted at rest. The plaintext .env repo became SOPS + age with three machines as recipients, every file round-trip verified before the plaintext was untracked.
• 143 branches → 2, every deleted branch preserved as a tag first. Recovery of any of them is one command.
• The audit was wrong twice, and that was the best part. The agent initially reported a repo had no backup and six repos didn't exist — both artifacts of checking GitHub with the wrong account. Because every claim carried its evidence, the corrections took minutes, and the real finding (my AI assistant's GitHub account could only see half the org) was more useful than the false ones.

What I think this means

The interesting part wasn't speed. It was the shape of the collaboration: the agent did the sweeps, the merges, the verification; I made maybe twenty decisions all day, each framed with evidence and a recommendation. The two-phase gate — audit first, act after approval — is what made the speed safe. And the standard exists so that the next agent session starts oriented instead of spending an hour rediscovering how everything fits.

More notes like this to come — this site is where I'm writing about building with AI: what works, what breaks, and what it changes about how one person can operate.