One system beats a swarm

Why multi-agent setups fall over, and what works better.

Ask the questionCan one AI take the busywork off my plate? Mostly, if you give it good notes.

You've probably seen the pitch: a team of AI agents, one for research, one for writing, one for sales, one to manage the others, all talking to each other while you sleep. It looks amazing in a demo. For a small business, in my experience, it's mostly rubbish. These setups break, they drift, and they give you different results every time you run them.

1. Open claude.ai. In the left menu choose Projects, then create a new project and name it after your business.
2. Open a plain document and write your standards file. Start with five headings: about the business, how we write, prices and offers, rules, and never do this.
3. Under each heading write a few honest lines. Short is fine. You'll add to it over time.
4. Back in the project, open the project instructions and paste: "Read the standards file before every task. If a request conflicts with it, tell me."
5. Add your standards file to the project's files, along with one or two key documents, like your price list or services list.
6. Start a chat inside the project and ask for something real, for example a reply to a customer enquiry.
7. Check the answer follows your standards. If it doesn't, don't just fix the reply. Add the missing rule to the standards file and upload the new version.
8. Create a second document called decisions log. Each time you make a call, add the date, what you decided and why.

Why the swarm falls over

• Every handoff loses something. Each agent only sees a summary of what the last one did. Detail and context fall out at every step, like a game of telephone.
• Errors compound. If one step is right nine times out of ten, a chain of five of them is right far less often. Small mistakes early become big mistakes at the end.
• Nobody owns the result. When the output is wrong, which agent got it wrong? You end up debugging a conversation between machines.
• It's hard to fix. Change one agent's instructions and three others behave differently. Simple problems turn into a project.
• It costs more. Every extra agent is extra usage, extra setup and another thing to maintain.

What works: one system, good notes

Here's how I actually work. One capable AI assistant does the job from start to finish, so it sees the whole picture. What makes it good isn't more agents. It's better notes.

• A standards file. One document that says how your business does things: your tone, your prices, your rules, your "never do this" list. The AI reads it at the start of every session.
• Project instructions. In claude.ai you can create a Project with its own instructions and files. Put your standards file and your key documents there, so every chat in that project starts with them.
• A decisions log. A running list of what you decided and why. It stops you, and the AI, going back over the same ground every week.
• One place for everything. Not notes split across five apps. One folder or one notes app the AI can be given.

When something goes wrong, you fix it once, in the notes, and it stays fixed. That's the whole trick. It's boring and it works.

How I actually run it

This isn't theory. It's how I run Media Canberra every day, with one AI assistant and one notes app:

• A standards file that loads every session. It lives in my notes app and the assistant loads it automatically before every job. When I catch myself repeating a rule, it goes in there once and applies everywhere.
• A memory index. One page, one line per note, so the assistant can see everything it knows at a glance and open only what it needs.
• One fact per note. Small notes stay true. When one turns out to be wrong, it gets fixed or deleted.
• Pattern notes. Anything I've built twice gets a note pointing at the best version, so the third time we copy it instead of starting over.
• A job number and a folder for every job. One register, one folder per job. Nothing gets lost.
• A wrap-up at the end of every session. What changed, what's live, what's left and why, written into the notes. Tomorrow starts warm.

How I brief it. My real requests are short and plain, and most are dictated by voice, typos and all: "working on [job], show me", "don't push until I say", "that's enough for today, add to memory". Name the job, set the limits, say how you'll check it. You don't need perfect prompts. You need clear ones.

Help me write a standards file for my business that an AI assistant will read before every task.

About my business: [what you do, where, who for]
How I like to write: [tone, words I use, words I hate]
Prices and offers: [what you charge, any rules about discounts]
Things that matter to me: [quality, speed, honesty, local, etc.]
Things I never want done: [e.g. promise dates I can't keep, use emojis, mention competitors]

Ask me up to ten questions to fill any gaps, one at a time. Then write the file with clear headings and short bullet points. Keep it under two pages. Plain English.

Email your AI like you'd email a staff member

This is one of the most useful things I've set up, and almost nobody does it. I talk to my AI assistant by email. Gmail has a feature called plus addressing: anything you add after a plus sign still lands in your own inbox. So yourname+claude@gmail.com is a second address that costs nothing and goes straight to you. That's the address I email my AI on.

What do I send it? Pretty much anything I'd hand to a staff member:

• Bug reports while I test, a screenshot or a phone video of something not working, with a line saying what I expected.
• Changes I want made, "move the phone number to the top of this page", with a photo of the page.
• Ideas, the moment I have them, so they're waiting for me at my desk instead of forgotten.
• Things from clients, forwarded straight in: their photos, their feedback, their wish list.

Here's why it's so handy:

• From anywhere. On a job, in a car park, on the couch. If I spot a problem on a website, I take a photo or a short video on my phone and email it in with one line. No laptop needed.
• Pictures and videos go straight in. A phone photo of a broken page explains more than a paragraph of typing ever will.
• Nothing gets lost. Every request sits in one labelled folder in Gmail, with a date, instead of scattered across chats, texts and sticky notes.
• Other people can report problems too. A business partner or a client can email a bug report to the same address, and it lands in the same queue.
• Picking it up is one sentence. When I sit down at my computer I say "check the latest email", and my assistant reads it, looks at the photos and gets on with it.

How the setup works

1. Pick the address: your Gmail with +claude added, for example yourname+claude@gmail.com. Nothing to create.
2. Add a Gmail filter so anything sent to that address skips your inbox and gets its own label, for example "Claude".
3. Give your assistant read-only access to that label only. I use Claude Code on my computer with a small script that reads the label, saves each email and its attachments to a folder, and never changes or deletes anything in Gmail.
4. Only your own addresses count. Anyone can email any address, so the script only accepts mail from addresses you've approved and that Gmail has verified as genuine. Everything else is ignored.
5. Say "check the latest email". That's it. The request is treated the same as if you'd typed it.

It works the other way too, with one rule: it only emails the updates someone actually asked for. "Email me when the fix is live." "Send the client their progress note on Friday." Not a running commentary. Keep a short written list of who asked for which update, tick each one off when it's sent, and check anything going to a client before it goes.