Build your own AI marketing and sales team with Claude and IdeaSpaces

1. The idea in one page

Most people start with prompts. They type a request, get a decent answer, and start again tomorrow from zero. The AI does not know their company, their customers, their prices or their voice, and it cannot touch their tools.

What works better is to build a small system once:

• Write down who you are and who you are not, in plain files the AI reads every time.
• Give the AI roles, one agent per job, each with clear limits.
• Connect the tools you already use, so agents can read and draft in them.
• Turn repeat work into written recipes (skills) and scheduled routines.
• Stay the editor. Agents draft, you approve. Nothing leaves the building without you.

I run my own marketing, sales and executive-assistant work this way. Today that is a dozen-plus agents and more than 40 written skills, all sitting on one laptop. This guide gives you the shape of it, without my specifics, so you can build yours.

2. The six layers

The order matters. Each layer depends on the one below it. If you skip the bottom, the top gives generic results.

The brain is plain text (Markdown) files in a folder on your laptop, backed up with Git, so you own them and they work with any AI model. I use IdeaSpaces to structure and back it up. Connectors (MCP connectors) let the AI read and act in your tools, so it works where your work already lives. Skills are written, step-by-step procedures the AI follows the same way each time. Routines are skills that run on a schedule or on demand.

3. What becomes possible

Think in job families, not tools. This is what a well-fed setup can do for a small team. Everything below stays a draft until a person approves it.

4. Write down what you are, and what you are not

This is the most important part, and it is the part people skip. The knowledge base is where the AI learns your business. Cover these:

Do not type it all. Talk. Dictate a rambling five-minute explanation of your business, then ask Claude to interview you and turn the answers into notes. Correct what is wrong. A first version in two hours beats a perfect version that never starts.

5. What should be in place before you start

• Claude on a plan that includes Claude Code or the desktop app, because agents need to read files, use tools and run tasks.
• IdeaSpaces (the plugin that gives the folder its structure and Git backup) and an empty folder on your laptop for your knowledge base.
• A backup remote (a private repository), so the brain survives a lost laptop.
• Two or three tools you already use and can connect: for example your email, your calendar and one more (CRM or website).
• A password manager and a rule: no passwords, ID numbers or bank details in notes.
• An always-on computer (optional), if you want routines to run overnight or want to control it from your phone.
• Two to three hours for the first setup, then about 30 minutes a week of tuning.

6. How to start: seven steps

1. Step 1. Set up the space · 30 minutes

   Install Claude and the IdeaSpaces plugin, create a folder for your knowledge base, and connect the backup.

   Done when: You open Claude in that folder and it can read and write a first note.

2. Step 2. Write the foundation · 60 to 90 minutes

   Use the checklist above. Speak, let Claude interview you, and keep each topic in its own note: company, ICP, offer, what you do not do, prices, team.

   Done when: A stranger could explain your business after reading it.

3. Step 3. Capture your brand and your voice · 60 minutes

   Gather your logo, colours and fonts, plus five samples of your best writing. Ask Claude to draft brand rules and a tone-of-voice guide, then edit them hard: add words you use, words you never use, and examples of good and bad.

   Done when: A new draft sounds like you without you rewriting it.

4. Step 4. Create your first agent · 30 minutes

   Pick the one job that eats the most of your time, usually inbox and calendar, or outreach. Give the agent its own folder and an agreement: what it does alone, what it asks you first, and what it never does.

   Done when: The agent can explain its own job and limits back to you.

5. Step 5. Connect two or three tools · 30 to 60 minutes

   Connect email and calendar first, then one more. Test read-only before you allow drafting or writing.

   Done when: The agent can answer “what is in my inbox and what is on my calendar today?” correctly.

6. Step 6. Write your first skill · 45 minutes

   Turn the job into steps: where to look, what to ignore, how to decide, what to draft, what never to send. Run it three times and fix the skill after every miss. A correction that is not written down will happen again.

   Done when: Three runs in a row need no more than small edits.

7. Step 7. Add a routine, then grow · 30 minutes

   Schedule something small and useful, such as a weekly summary. Add a new agent only when a job repeats and the first agent is working well. Every new agent gets its own folder, agreement and skills.

   Done when: One task runs by itself and you only review the result.

7. Rules that keep it safe and useful

• Agents draft, you approve. Emails, posts, invoices and payments leave only after you say so.
• Keep secrets out. No passwords or personal ID numbers in notes. Keep private matters in a separate space the work agents cannot read.
• Check numbers against the source on the day. Notes go stale. Put a date on every fact.
• Write down every correction. After a mistake, update the skill or the note. This is how the system gets better each week.
• Respect data rules. Do not paste customer personal data into tools your company has not approved.
• Grow slowly. One agent working well is worth more than five half-built ones.

8. Common mistakes

9. What it looks like after a few months

The biggest wins for me were not the flashy ones. The inbox gets checked and sorted across several accounts. Follow-ups happen on time. Every call ends with a written list of who does what. Outreach is researched and personalised, in small batches I can control. Reports use real numbers instead of last month's memory.

What still needs me: strategy, relationships, pricing, and the final read of anything that goes out. The interface is plain, because you work in a chat window, and it needs a little care every week. In return, the knowledge stays in files you own and works with whichever AI model you choose next year.