Put it on autopilot (almost).
Agents are the hot topic and the muddiest word in AI. This lesson de-hypes it, shows you the automation ladder, and gets your first repeatable workflow running with the right guardrails.
Lesson 07 of 07 · ~12 min · FreeWhat an agent actually is
A chat answers a question. An agent is Claude given a goal, some tools and permission to take multiple steps toward that goal: read this, then draft that, then check it, then format it, without you prompting each step. That's the whole definition. Everything else is marketing.
You've already met the ingredients. Projects gave Claude standing context (Lesson 3). Connectors gave it hands (Lesson 4). An agent is those two plus a repeatable set of steps. Which means you don't start by "building an agent", you start by designing a workflow you'd trust.
The automation ladder
Four rungs. Each adds leverage, and each needs a bit more governance. Most marketers get enormous value from rungs one to three.
Automate the mechanical, gate the voice
Here's the design principle that separates useful automation from generic slop. The mechanical steps (gathering, formatting, resizing, scheduling, summarising) can be fully automated: nobody's brand died from an auto-formatted spreadsheet. Anything carrying your voice or judgement (the actual words, the strategy, the send button) keeps a human gate. Claude pre-fills, you approve.
And a second rule from hard experience: never automate something you haven't done manually at least three times. Automation doesn't fix a process you don't understand, it just produces the mess faster.
Design your first workflow
Pick one job you do every week that follows the same shape. Repurposing content, a weekly report, tidying a data export. Then run this.
- Write down one job you do every week that follows the same steps each time.
- Check it passes the three-manual-runs test: you've done it by hand often enough to know what good looks like.
- Open (or create) the relevant Project from Lesson 3, then run the workflow designer below inside it.
The workflow designer
Run inside the relevant Project. Keep the result as your master prompt.
Here's a job I do every week: [describe it, step by step, as you do it manually]. Design a repeatable Claude workflow for it: 1. The steps, in order, and which ones you can do vs which need me 2. What context should live in this Project so I never re-explain it 3. Where the human approval gates belong (anything with my voice, or that sends) 4. The single master prompt I should run each time Then ask me the questions you need answered to make it work first time.
The pre-automation check
Run before climbing to rung four with any job.
I'm considering automating this job: [describe it]. Play sceptic. What could go wrong if this ran without me watching? What's the worst output that could ship? Which steps genuinely need my judgement? Tell me honestly whether this is ready to automate or needs more manual runs first.
- The agent landscape moves monthly. Right now Claude's agentic tools include Claude Code (for developers) and desktop tools that run multi-step tasks with your files and apps: check claude.com for the current line-up before choosing one.
- Start on rungs one to three in the chat you already know. Climb to rung four only for a workflow you've already run successfully by hand.
- Whatever the tool, the principle doesn't move: mechanical steps automate, voice and sends stay gated.
Automating a bad process.
Automation is an amplifier. Point it at a process that works and you multiply the value; point it at a mess and you industrialise the mess. Three manual runs first, always.
Three questions, then the good bit
No marks, no sign-up, just a check the lesson stuck. Tap your answer.
The difference between a chat and an agent is:
The first thing worth automating is:
The rule that keeps automation from going generic:
Want your whole business working like this?
Brandling helps small businesses put AI to work without losing their voice. Senior help, right-sized.
Book a free triage →