# Build Your Own Agent Team
### A starter kit for solo consultants and independent professionals

You are running a whole business by yourself: finding the next client, staying visible, delivering the work, sending the invoices. Most of that is coordination and preparation, not judgment. This kit shows you how to hand the coordination and preparation to a small "team" of Claude agents while you keep every decision that matters.

No special software required. If you can use Claude in a browser, you can build this.

---

## 1. What an "agent" actually is

An agent is just a written set of instructions telling Claude who to be for one specific job. That's it. You write a one-page job description, give it to Claude, and Claude plays that role consistently: same priorities, same boundaries, same output format, every time.

The magic is not in any one agent. It is in splitting your business into narrow roles so each conversation has one clear job, instead of one giant chat that tries to do everything and forgets what matters.

## 2. The team structure

You build a hub and spokes:

- One **coordinator** (think chief of staff). This is the only agent you talk to day to day. It routes work, chases loose ends, and gives you a short daily or weekly brief organized around what needs YOUR attention.
- A few **specialists**, each owning one narrow function of your business.

For a solo consulting practice, the specialist roles that earn their keep are:

- **Pipeline** finds and helps you win the next piece of work: tracking leads, drafting outreach, watching for opportunities, keeping a simple pipeline list current.
- **Marketing** keeps you visible: drafting posts, newsletter issues, talk proposals, and repurposing your delivered work into public material.
- **Research** supports delivery: background briefs before client meetings, industry scans, summaries of reports you don't have time to read.
- **Ops & Finance** gets you paid and keeps the back office moving: invoice tracking, follow-ups on receivables, contract checklists, expense summaries.

Do not build all five on day one. Start with the coordinator plus the one specialist that hurts most (for most consultants, that's Pipeline). Add the others only when you feel the gap.

## 3. Escalation barriers: the rule that makes this safe

This is the most important section in this guide. Every agent needs a hard, written line between what it may do alone and what it must bring to you. Without that line, you either get an agent too timid to be useful or one confident enough to be dangerous.

Write the barrier into every agent's instructions as two explicit lists:

**Always escalate (the agent must ask, never act):**

- Taking, declining, or pricing any piece of work
- Anything that commits your name, time, money, or reputation externally
- Sending anything to a client, prospect, or partner that isn't already an approved pattern
- Trade-offs between clients or between lines of business
- Anything novel, ambiguous, or higher-stakes than the established routine

**May decide alone (the agent should just do it):**

- Drafting, researching, organizing, and formatting
- Sequencing and scheduling its own work
- Routine follow-ups that match a pattern you've already approved
- Routing a task to the right specialist (coordinator only)

The test that resolves every edge case: **if it would change what you think, owe, or own, the agent surfaces it and asks. If it just moves existing work forward, the agent does it.**

Two habits keep the barrier working over time. First, when an agent escalates something you consider routine, don't just answer; update its instructions so that pattern is now pre-approved. The barrier is meant to move as trust builds. Second, when an agent quietly does something you wish it had asked about, tighten the written line the same day. The instructions file is the contract; keep it current.

## 4. How to build it (no paid software needed)

You can run the whole team in regular Claude chat using **Projects**:

1. Create one Project per agent (Projects are available on the free plan, up to five, which is exactly a coordinator plus four specialists).
2. Fill in one copy of the TEMPLATE file for each agent.
3. Paste that agent's completed instructions into its Project's custom instructions.
4. Talk to each agent inside its own Project.

The one limitation of this setup: the agents can't talk to each other automatically, so you are the messenger. In practice this is quick. The coordinator says "this is a Pipeline job, here's the brief," and you paste the brief into the Pipeline Project. The round trip takes a minute, and you see everything that moves between agents, which early on is a feature rather than a cost.

If you later move to a tool where agents can invoke each other, the same instruction files carry over unchanged. The roles and the escalation barriers are the design; the software is just plumbing. (When you're ready for a more integrated setup, see "Level 2" below.)

## 5. Working rhythm

- **Daily (5 minutes):** open the coordinator, ask for the daily brief. It should come back in three sections: **Needs you** (decisions only you can make), **Moving** (work in flight), **FYI**. If nothing needs you, a good coordinator says exactly that.
- **Weekly (30 minutes):** ask the coordinator for a weekly synthesis: pipeline health, upcoming commitments, pending decisions, anything slipping. Feed each specialist any updates from your week they couldn't see.
- **As things happen:** forward new information to the relevant specialist Project so briefs stay current.

## 6. Level 2: the same agents as "skills" (optional, skip on day one)

The Projects setup above is the right place to start. But Claude also supports **skills**: named roles you install once and then summon inside any conversation, instead of keeping one Project per agent. Where a Project is a separate room you walk into, a skill is a colleague you can call into whatever room you're already in.

Two extra files in this kit are the same Sterling and Marlowe repackaged in skill format (SKILL-VERSION-1 and SKILL-VERSION-2). The content is identical: same role, same escalation barrier, same outputs. What's different is the packaging:

- A short block at the top (the lines between the `---` marks) gives the skill a **name** and a **description**. The description matters: it's how Claude knows when this role is the right one to adopt, so if you rename the agent or change its job, update the description to match.
- The instructions are written to take over the conversation when invoked ("act as Sterling for the rest of this conversation"), rather than to define a whole Project.
- The business-context section sits at the bottom as a fill-in block. Complete it before first use, same as the template.

How to install them, depending on where you run Claude:

- **In the Claude apps (paid plans):** look for Skills in Settings and upload the file there, or ask Claude itself to "create a skill from this file" and paste the contents.
- **In Claude Code (the developer tool):** each skill lives in its own folder as a file named `SKILL.md`. Save SKILL-VERSION-1 as `~/.claude/skills/sterling/SKILL.md`, and likewise for Marlowe.

### Summoning an agent with a slash command

Once a skill is installed, its name doubles as a **slash command**: a shortcut you type at the start of a message to call that agent into the conversation.

How to set it up: you already did. The `name:` line in the skill file's top block IS the slash command. `name: sterling` means typing `/sterling` summons Sterling. If you rename your agent, change that line; keep it lowercase with no spaces (hyphens are fine, so `name: my-chief-of-staff` gives you `/my-chief-of-staff`).

How it works in practice:

- Type `/` at the start of a message and your installed skills appear in a list; keep typing to narrow it down.
- `/sterling` alone loads Sterling's instructions, and Claude acts as Sterling for the rest of the conversation.
- You can hand over an assignment in the same breath: `/sterling weekly synthesis` or `/marlowe brief me on the Acme introduction`. Whatever follows the command becomes the agent's first task.
- To bring in a different agent mid-conversation, just call its command: get your brief from `/sterling`, then type `/marlowe` in the same chat to go deeper on pipeline. Both agents can see the whole conversation, so nothing needs re-explaining. That shared memory is the main payoff over the Projects setup, where you are the messenger between separate rooms.

What's actually happening under the hood, so none of this feels like magic: a slash command simply pastes that skill's instruction file into the conversation for you. It is exactly the same as copying the file in by hand, reduced to one keystroke. Which is also the reassurance: if slash commands don't appear where you're using Claude, you can always say "act according to the sterling skill" or paste the file's contents directly, and the agent behaves identically.

One extra option if you use Claude Code: it also supports bare custom commands with no skill packaging at all. A plain text file saved as `~/.claude/commands/sterling.md` becomes `/sterling`, and its contents are sent as the prompt when you invoke it (with `$ARGUMENTS` standing in for anything you typed after the command). It's a fine lightweight alternative, but for agents I recommend the skill format above: the description line lets Claude suggest the right agent even when you forget to summon it.

The cost of the skill setup is a little more installation work and a paid plan. The escalation barriers work exactly the same in both formats, and that is the part that matters.

## 7. Level 3: put the daily brief on a schedule (optional)

Everything so far runs when you show up. The last upgrade is making the coordinator run before you do, so the daily brief is waiting with your coffee instead of being something you remember to request.

Claude's Cowork feature (in the paid plans) supports **scheduled tasks**: you write a prompt once, pick a cadence (daily, weekdays, weekly, hourly), and Claude runs it on schedule even while your computer is asleep, with the result waiting when you check in. Look for "Scheduled" in the Cowork sidebar, create a new task, and either describe what you want in plain language or set it up manually.

Two things to know when you write a scheduled prompt:

- **Each run starts fresh, with no memory of yesterday.** So the prompt has to carry everything: point it at your coordinator's instructions (or its skill) and spell out the job. Something like: "Follow the attached Sterling instructions. Produce the daily brief: Needs you / Moving / FYI, one screen." A vague prompt that works fine when you're there to clarify will drift when nobody is watching.
- **Escalation barriers matter more on a schedule, not less.** A scheduled run happens while you are away, which is exactly when a fuzzy boundary becomes a problem. The rule for unattended runs is simple: they read, prepare, and draft; they never send. Anything on the "always escalate" list comes back as a question in the brief, where it waits for you.

If scheduled tasks aren't available where you use Claude, a recurring calendar reminder plus a saved message you paste each morning gets you most of the benefit. The ritual is the point; the automation is a convenience.

## 8. What's in this kit

- **This guide.** The approach and the rules.
- **TEMPLATE-Agent-Instructions-Fill-in-the-Blank.** Copy it once per agent and fill in the blanks. The escalation section is not optional.
- **SAMPLE-1-Coordinator.** A finished coordinator instruction file for a solo consulting practice, so you can see the template filled in.
- **SAMPLE-2-Specialist-Pipeline.** A finished specialist example for the Pipeline role.
- **SKILL-VERSION-1-Coordinator** and **SKILL-VERSION-2-Specialist-Pipeline.** The same two agents in skill format, for when you outgrow the Projects setup (see Level 2 above).

Start with the coordinator today. Fill in the template, paste it into a Project, and ask it for your first daily brief. You will know within a week which specialist to build next.


---

*Part of the Build Your Own Agent Team starter kit by Kimberly Bella. The kit, and the story of the agent team it came from, live at [plaininference.com/agent-team-kit](https://plaininference.com/agent-team-kit).*