# Set up your first AI agent

**The CV Method · The Playbook · No. 2 · Working with AI · Aug 2026**
*by Corina · @the_cv_method · hello.thecvmethod.com · current as of Aug 2026 — verify tool details, this space moves fast*

A plain-English guide to standing up a small, reliable AI "worker" that does one job well and never goes rogue — using tools you already have. This is how I actually set one up, minus the hype.

---

## First, the mindset

An "agent" sounds mystical. It isn't. It's an AI model given one clear job, the right context, a few tools it's allowed to touch, and guardrails — running with a human in the loop. Small and specific beats big and magical every time. Everything below rests on one rule.

## Rule zero — it proposes, you approve

Your first agent should never post, send, spend, or deploy on its own. It drafts and suggests; a human ships. That single rule is what makes an agent safe to actually use.

## Step 1 · Pick your stack

Four simple pieces. Here's what I use and the general equivalent — swap in whatever you already have.

| Piece | What it does | I use | Swap for |
|---|---|---|---|
| **The brain** | An AI model that can follow instructions and call tools. | Claude | Any strong tool-use LLM |
| **The runner** | Where the agent actually runs and takes its actions. | Claude Code / an agent runner | A simple script · a workflow tool |
| **Its tools** | The connectors it's allowed to touch (files, calendar, a doc, an API). | MCP connectors | Plugins · a few scoped API keys |
| **Its memory** | A place for its instructions + the files it reads and writes. | A notes vault (files) | Plain folders · a doc |

Give it the fewest tools that get the job done. Every extra permission is a new way for it to surprise you.

## Step 2 · Give it ONE job

Don't build "an assistant that does everything." Build "the thing that drafts my weekly update," or "the thing that reads a receipt into a row." Write the job as a short brief: its role, the goal, the constraints, and what "done" looks like.

If you can't describe the job in two sentences, it's still too big. Split it.

## Step 3 · Give it the right context

An agent is only as good as what it can see. Hand it the specific docs, examples, and data the job needs (a style doc, last week's update, the folder of receipts) — not your entire world. Curated context makes it reliable; a firehose makes it wander.

Keep a short "how we do this" doc the agent reads every run. That doc is where most of your quality lives.

## Step 4 · Set the guardrails

Spell out the limits in its instructions: it drafts, never publishes; it asks before anything irreversible; it stays in its lane (only the tools and files it needs). Give it the least access that works, and a clear "when in doubt, stop and flag me."

The goal isn't an agent that can do anything. It's one you can trust with a small, real thing.

## Step 5 · Test before you trust it

Before you automate anything, run the agent manually on a handful of real inputs. Read every output. Where it's wrong, that's not failure — that's the instruction or the context telling you what to tighten. Repeat until it's boringly consistent.

Only after it's reliable by hand should you put it on a schedule — and even then, add a fail-safe so it flags you instead of failing silently. It still never auto-posts.

## Keep it honest & safe

- You approve before anything ships.
- Least access — only the tools it needs.
- No invented results — read the outputs.
- Keep keys & secrets out of prompts/logs.
- Respect the terms of anything it touches.
- A human owns the irreversible actions.

## Steal these prompts

**The one-job brief — fill in and hand to your agent**

```
You are my [role] agent. Your ONE job: [the single task — e.g. "draft my weekly update from these notes"].
Context you may use: [the specific docs/files — not everything].
Guardrails: you draft only — never send, post, spend, or publish; ask before anything irreversible;
use only [these tools]; if you're unsure, stop and flag me.
"Done" looks like: [describe the finished output].
Do the task now on this input: [paste the real input].
```

## The starter checklist

1. Write Rule zero: it proposes, you approve.
2. Pick the 4 pieces (brain, runner, tools, memory).
3. Name ONE job in two sentences.
4. Write the brief: role, goal, constraints, "done."
5. Give it the specific context — a "how we do this" doc.
6. Set guardrails + least access.
7. Run it by hand on real cases; tighten.
8. Only then schedule it — with a fail-safe, never auto-posting.

---

**Want a fleet of these, built for your business?** One reliable agent is a great start. A set of them running your back office — safely, with the guardrails — is the thing I build for people. Start at **hello.thecvmethod.com**.

*The CV Method · Set up your first AI agent · No. 2*