I run my company's ops through a fleet of Claude Code bots. Here's what 30+ rebuilds taught me.

We're a deliberately bootstrapped startup — group of founders, mostly GTM-leaning, no immediate plans to raise. Which left us with a problem: how do we deliver a full C-suite's worth of functions without hiring a single expensive exec? Eight weeks ago I set out to answer the follow-up — could I build that missing C-suite in Claude Code?

Not a chatbot bolted on the side — actual teammates. Software shaped around a role, not a feature. Here's where it's landed, including the bits that went sideways.

What's actually in production:

A CMO function the whole team uses daily. First one I built, and by far the hardest — 30+ separate worktrees before I found something that held up in real use. Most were dead ends: too broad, too vague, confidently hallucinating campaign nonsense. The version that survived is narrow and opinionated (maybe a little too much!).

A COO function I lean on hard for the unglamorous core — team management, product development, GitHub management. Sounds boring. Turned out to be the one I would genuinely struggle to give up.

And the fleet runs first- and second-line customer support: it triages incoming feedback, finds the actual bug buried inside a vague complaint, and routes it to the dev team. That one workflow has saved more hours than anything else I've built. By a mile.

(There are lower-fidelity bots for finance, compliance, tax too — but those three are the ones earning their keep.)

Three things that genuinely surprised me:

1. The boring bot won. I expected the CMO to be the star. It's fine! But the COO — the one doing unsexy ops coordination — is the one I'd fight to keep. The value wasn't any clever output. It was having something hold the operational thread so I stop dropping things.

2. Support triage was the highest-ROI thing, by a distance. Turning vague customer moans into clean, routed bug tickets sounds mundane. It quietly killed off a whole category of work w used to do slowly and badly.

3. The hard part was two problems, not one — and I only saw the second one late. Scoping was the obvious half: every bot that failed, failed because it was too broad. "Be my CMO" produces confident nonsense; "own this exact slice, these inputs, these outputs" produces something the team trusts. But the half I badly underestimated was functionality — actually wiring the thing to execute the role, not just reason about it. Reliable Slack comms, the right access to the systems it needed (Drive, Slack, the marketing tools), and genuinely building up its skills so it could act rather than advise. A perfectly-scoped bot that can't reliably do the thing is still a broken product. Most of those 30+ rebuilds were me grinding through both problems at once.

Honest caveats: it took real time, it costs real money in tokens (the cost-control story is a whole separate post…), and a lot of what I built only makes sense for my company. The transferable thing is the pattern, not the code.

Why I'm posting this: the most fun I had wasn't building for my own company — it was building with other people. I taught my son to code by building a tool together (he's since taken it and run with it on his own setup — quietly proud-dad moment). And I helped my brother automate a chunk of work he'd dreaded for years — watching him actually use it, daily, is what hooked me. That's the part I'd happily do more of, so alongside the day job I do a bit of advisory for founders building their own role-shaped bots. Not a pitch — if that's you, just say hi.

The thing I'm actually curious about: for those who've tried turning a human function into a bot — which function did you pick, and where did it break? Trying to map where this pattern holds and where it falls apart.