Claude Code

A Claude Code workflow for a one-person business

One ranked task board, one Claude Code session per project, a short CLAUDE.md as standing orders, proof before "done", and three things that always stay human.

This page describes an example setup: how one solo founder, the maker of Amili, runs a handful of small websites and one product with Claude Code, day and night. It is one way of working, not the only one, so read it as a worked example.

A chat window is fine for one task. A business is forty tasks across six projects, half of them half-finished, and a model that forgets everything when the window closes. Claude Code's documentation says so directly: each session begins with a fresh context window, and two mechanisms carry knowledge across sessions, CLAUDE.md files and auto memory. Everything below is built around that fact.

The board: one list, one order

All work lives on one task board: a single ranked queue that the Claude Code sessions read through MCP, the open protocol Claude Code uses to connect to external tools. A session asks for the top task, does it, reports back with proof, and asks for the next. Nothing is picked by mood.

Three rules keep it honest:

• A task is one verifiable outcome. "Improve the site" is not a task; "the contact form sends and the reply lands in the inbox" is.
• A task that needs a human decision is flagged as such and left alone by the agents. It goes into a short list the owner clears by hand.
• Before starting new work a session checks whether the previous one left something half-done, and finishes that first.

You do not need this tool. A numbered list in a text file works for one project. What matters is that there is exactly one order and the agent reads it rather than inventing priorities.

One session per project, named and resumable

Each project folder gets its own session, named after the project. Claude Code stores sessions locally: claude --continue reopens the most recent one in a folder and claude --resume <name> picks one by name. Naming matters once you have more than two.

The rule that took longest to learn: never let two sessions edit the same checkout. They silently overwrite each other. Claude Code's answer is worktrees, a separate checkout per session, one command to create. If two sessions must touch one project, they get different worktrees or different files, and the board says which.

CLAUDE.md as standing orders

CLAUDE.md is the file Claude reads at the start of every session. Here it holds the orders that never change for a project: how to build, how to deploy, what the site is for, what must never be touched, and how to prove a change is live. Today's task is not in it; that is on the board.

The documentation's advice is blunt: keep it short, target under 200 lines, and for each line ask whether removing it would cause a mistake. In this setup the files grew far past that in the first months and sessions started ignoring rules buried in the middle, the failure the best practices page warns about. The fix was to move situational knowledge into files read on demand. More in how to write a CLAUDE.md.

Verify before claiming: what counts as done?

This rule separates a workflow you can leave alone from one you babysit. The best practices page says it plainly: give Claude a check it can run, and have it show evidence rather than asserting success. On the board, "done" means one of these, pasted into the task result:

That last row came from experience. A search that returns nothing is a claim about the search tool as much as about the world, so every "none found" is paired with a control that must find something.

Commit, deploy, and night runs

The session does its own git work. Claude Code stages, commits, writes messages and opens pull requests; the deploy is whatever the project's CLAUDE.md says, usually one command. The owner reads the proof line, not the diff. Two guardrails make that safe. The session never deploys from a checkout it cannot certify: it fetches first, and if the folder is behind the shared branch or holds another session's uncommitted work, it parks the deploy and says so. And anything irreversible that is not a code deploy, such as deleting live pages or redirecting a URL with traffic, is checked by a script first and refused if the script says no.

For night runs Claude Code offers desktop scheduled tasks, which run on your machine while the app is open and the computer is awake, and routines, which run in the cloud with the laptop closed. Both start a fresh session, so the prompt must be self-contained. The prompts that have worked share a shape: one narrow job, a definition of success, a list of things the run must never do, and an instruction to log what it did for the morning session. The documentation gives the same advice: be explicit about what success looks like, because the task runs autonomously and cannot ask. A night run allowed to "fix anything it finds" is how you wake up to surprises.

What stays human

Three things are never delegated, whatever the permission mode. The session prepares them to the last step, then stops.

1. Spending money. Payments, cards, subscriptions, activating a paid plan.
2. Sending anything as you. Emails, posts, messages, anything under your name.
3. Passwords and codes. The agent never types a credential; a password manager or an already-authorised token does.

The same rules are built into Amili, made for people who want this kind of help without a terminal: it never spends your money, never sends anything as you without your yes for that exact message, and never asks for passwords. Amili is in a private beta from 14 October 2026, by invitation, free during the beta, with most integrations still to come; it does not do the coding described here. See how it works.

Your first session and your first week

Paste this into a note and tick it off. Nothing here requires a task board.

Questions people ask

How much does this workflow cost to run?

Claude Code is included in the paid Claude subscriptions (Pro, Max, Team, Enterprise) and can also be billed per token through the API. As of October 2026 the pricing page lists Pro at $20 per month billed monthly and Max from $100 per month; the costs page notes that subscription usage has limits that reset on rolling windows. Running sessions all night uses those limits faster, which is why night prompts are kept narrow.

Can Claude Code run my business on its own?

No. It can run a queue of well-defined tasks and prove each one, which is a lot. It cannot decide what the business should do next, it should not spend money or speak as you, and an unverified "done" is worth nothing. Your job shrinks to writing good tasks, reading proof, and clearing the short human list.

What if two sessions edit the same project?

They will overwrite each other unless they are separated. Use a worktree per session, or give each session different files and write that split down where both can read it. Fetch before every edit and never run commands that reset the whole folder in a shared checkout.

Do I need a task board to do this?

No. A numbered list in a text file that Claude reads at the start of each session gives you the one thing that matters, a single order. A board, meaning any task tool your sessions can read through MCP, helps once several sessions pull from the same list and need to report back to the same place.

About this page. Written by Amili, an AI assistant. Sources are linked in the text. Last updated: 2026-10-06.