Inside a 1-operator AI studio: the actual workflow

The studio shipped four websites, two CRMs, an email automation rebuild, and a brand refresh last month. It is one person and a permanent team of agents. People keep asking how the math works, so here is the unromantic version.

The day starts in a triage queue

Every morning, an agent sweeps Slack, email, GitHub, and Linear and produces a single page: what's blocked, what's overdue, what's quietly broken in production, and what a client said yesterday that I haven't replied to. It takes me twelve minutes to read it. Without it, I'd lose two hours every morning playing detective.

The agent doesn't decide priorities. It surfaces them. I still pick the order. The leverage is on the surface, not the decision.

Build work runs in two modes

I run two patterns depending on the certainty of the spec. When I know exactly what to build — a new section, a CRM field, a Stripe webhook — I delegate to a coding agent with a tight brief and watch the diff. The agent does the typing; I do the review. When the spec is fuzzy, I drive directly with the model alongside me, talking out loud about constraints, and we shape the spec together before any code lands. The fuzzy mode is slower but it's the only one that doesn't ship the wrong thing in beautiful syntax.

The mistake new operators make is using delegation mode for fuzzy work. The agent can't ask the questions you didn't think to answer. If you're not sure what you want, sit with the model, not above it.

The four checkpoints I never skip

• Before any client call: a fifteen-minute prep where the model reads the project history and writes me a one-page state summary plus three questions I might have missed.
• Before any deploy: a checklist run by the agent — tests pass, types clean, no console errors in preview, lighthouse over 90, copy reviewed. If any item is red, the deploy is blocked.
• Before any invoice goes out: a reconciliation pass that compares scope-of-work to what shipped. Twice this year it caught me undercharging by hundreds because I forgot what was added mid-project.
• End of week: an honest retrospective with the model. What took longer than expected, what was easy, what should be a template next time. I keep these. They become the most valuable thing in the studio.

What I never delegate

Three things stay on me, always: the first conversation with a new client, the visual taste calls in design, and the moment I tell someone I can't help them. Agents can draft, summarize, route. They can't represent the studio. The studio is the relationship, and the relationship lives in the messy human seams a model can't cross.

The tools, named

• Claude Sonnet 4.6 — daily driver. Reads code, writes code, drafts copy, runs the morning brief.
• Claude Code + Cursor — the actual building. MCP-connected to Figma, Postgres, GitHub, and the n8n queue.
• n8n (self-hosted) — every recurring trigger lives here. Webhooks, cron, retries, dead-letter handling. Integrations service →
• Linear — single source of truth for what's promised and what's done. The triage agent writes here.
• Stripe + Resend — money in, mail out. Both wired through n8n with monitoring on every webhook.

The economics

AI inference is now the second-largest line item in the studio after taxes. It also pays for itself by a factor of roughly twelve. That ratio is not a forecast — it's the actual number from this year's books. If you are running a small services business and your AI spend isn't visible on the P&L, you aren't using it hard enough. Want this set up for your team? See AI enablement.

The promise of one-person studios isn't "replace your team with bots." It's that the floor of what one careful operator can deliver moved by an order of magnitude in three years. Most people are still working as if the floor stayed where it was.