From talking to AI, to AI that does the work.
Last time we learned how to talk to AI. Today we hand it the job.
Last time you learned to talk to AI. Today it starts doing the work.
On 20th May we covered the foundations. What an AI model is, how to prompt it, and how to keep your data safe. You can now hold a good conversation with AI. That was step one.
Session 1 (done)
- Ask it a question.
- Get an answer.
- You drive every word.
- Good for one-off help.
Session 2 (today)
- Hand it a task.
- It does the steps.
- You check the result.
- Good for work you repeat.
Here is the real reason this matters, and it is a true story from my own month.
Every month I have to deal with invoices. I used to do it by hand. Now I hand the pile to Claude. It sorts them, builds an expense spreadsheet, tracks where each invoice is in the process, and drafts the email to my accountant. I check the work, then it goes out. The same set of steps would run in OpenAI's tools or Google's tools. The question is no longer whether AI can do this. It can. The question is how you set it up to do it the same way every time. And all of that starts with one good prompt.
A good prompt is how you tell an agent exactly what you want.
We are moving from talking to AI toward handing it whole tasks. The prompt does not go away. It becomes the instruction you give the agent. A vague prompt gets you a vague worker. A clear one gets you a worker that does the job your way. Six parts make a prompt clear.
Who you want the AI to be
One line. "You are my bookkeeping assistant." That one line narrows and focuses what the model draws on. It points the model at the right slice of what it knows, so it answers as that role.
The one job, stated plainly
"Sort these invoices and log them." Spell the job out as a clear step-by-step list, not one vague ask.
What it needs to know
The background, the actual inputs, what you already tried, anything it would have to guess otherwise.
What you want it to answer back
Name the shape. A table, a short list, a draft email, three bullet points. Ask and you get it.
Show one example of right
So it copies the shape. One good sample beats a paragraph of description.
What it must not do
What to leave to you. "Draft only, do not send. Flag anything missing instead of guessing." This is where you stay in control.
A few vague lines. No role, no inputs, no format, no example, no limits. It looks reasonable, but the AI has to invent the half that matters.
Sort my invoices and put them in a spreadsheet. Then email my accountant about it.
All six parts, covering the whole job. Role, the full multi-step task, context, the format you want back, an example to match, and clear boundaries.
Role: you are a careful bookkeeping assistant. Task: get my May invoices ready for my accountant. Four steps. 1. Locate the invoices for the period. 2. Build an expense spreadsheet from them. 3. Check with me on anything unclear before you finish. 4. Then draft the email to my accountant that goes with the file. Context: the invoices are in my "Invoices/May" folder. The period is 1 to 31 May. I already have an expense sheet from last month, "Expenses 2026", so match its layout and add this month to it. Format: - Spreadsheet columns: date, vendor, description, amount, category. - Email: two short paragraphs, then a bullet list of anything that needs her attention. Example of the tone I want for the email: "Hi Dana, here is May. Numbers are in the file, two items need a second look." Boundaries: pause and ask me before you finish. Draft the email only, never send it. Never change the numbers on an invoice. If something looks missing or off, flag it for me instead of guessing.
A good prompt like that one is worth keeping. Save it, and you have written down how this task should be done.
The thorough prompt, written out as steps, is your SOP.
Take that thorough prompt and write out, in detail, how the task gets done. Each step, in order. That written-out version is your SOP. The prompt was the instruction. The SOP is the durable version, the thing you keep and reuse. SOP, playbook, workflow are three names for the same thing. A step-by-step guide for how you would do a task. Different tools call it different names. You break a big task into ordered steps, and where a step is its own small job, you give it its own small prompt under the main one. One blob of text does one thing badly. Ordered steps do the whole job.
A good SOP has the same shape every time, whatever you call it.
This is what a good SOP looks like. Every step has a clear path, an example, and a way to know it worked.
| Step | Path | Example | Success criteria |
|---|---|---|---|
| 1. Find the invoices | Open the May invoices folder, list each one | 12 PDFs, 1 to 31 May | Every invoice for the month captured, none missed |
| 2. Build a spreadsheet | One row per invoice, match last month's columns | Date, vendor, amount, category | Totals add up, no blank cells |
| 3. Check with me | List anything unclear, ask before going on | "Two invoices have no date" | I have answered before it drafts anything |
| 4. Draft the email | A short summary plus the attached sheet | "Hi Dana, here is May..." | Draft sitting in my drafts, not sent |
Guardrails are the lines your AI must never cross.
Guardrails matter most in agentic work, because the AI is acting on its own. The SOPs we built are how you put guardrails in place, so the AI follows the lines you set. You can also set general guardrails on the tool itself.
These are the general guardrails you set on the tool itself, the real switches behind the lines above. Three places to set them.
What the agent can see and reach
Which apps, which files, which accounts.
Permissions and approvals
What needs your sign-off, and what it can never do.
Tone, channel, and audience
Tone, which channel it uses, and who it is allowed to talk to.
Paste this at the end of any task where being wrong is expensive. It tells the model to grade its own confidence and admit the gaps instead of bluffing.
Before you give me the final answer, do a brutally honest review of your own work. 1. Rate your confidence in each part from low to high. 2. Flag anything you are not sure about, and say why. 3. If you do not actually know something, say "I do not know" instead of guessing. 4. List what a human should double-check before this goes out. I would rather have an honest "I am not sure" than a confident answer that is wrong.
Three words people mix up: assistant, workflow, agent.
A smart helper at your elbow
You ask, it answers, you decide the next move. You drive every step.
Example: "Rewrite this email to sound warmer."
A recipe that runs itself
The same way every time. Fixed steps, no judgment. You set it once.
Example: every new invoice gets renamed, filed, and logged the same way.
A junior employee you hand a goal
Not a script. It figures out the steps and does them. It needs supervision, like any new hire.
Example: "Get my invoices ready for the accountant."
Tools: examples are Gemini Spark, Claude Cowork, and ChatGPT agent.
A Gemini Gem, a Claude Project or artifact, a custom GPT. These feel advanced, but they are still assistants. Each one is a chat with a fixed bundle of knowledge. It gives you information, it does not go and do the task. Adding a notebook like NotebookLM on top only gives it more to read. Still an assistant, only better informed.
Match the tool type to the job. Take a common one. You want to pull customer reviews from a few sites, sort them, and write a short summary you can read on Monday.
Build it as workflows
- Break the job into eight to ten small steps.
- One pulls the reviews. One cleans them up. One sorts by rating. One writes the summary.
- Each step does one thing and never changes. You wire them in order.
- Best when the steps are fixed and you want it predictable.
Hand it to an agent
- Give one agent the whole goal.
- "Pull this week's reviews from these sites, find the themes, and write me a short summary."
- It works out the steps itself and does them together.
- Best when the job shifts week to week and needs some judgment.
Your tools today, and what is coming.
Your team already lives in Google, so here is what you can reach today and what is on the way. Same recipe, different stoves.
| What you want | Tool | Effort | Status |
|---|---|---|---|
| Gemini right inside your docs | Google Workspace | Zero, it is already there | Working now |
| Build an agent, no code | Google Workspace Studio | Plain English | In your plan now |
| Build an agent, the technical route | Google AI Studio | For developers | Now |
| The easy all-in-one assistant | Gemini Spark | Email it like a colleague | US beta, soon |
Gemini Spark is the easy version of all this. You email it at its own address like you would email a colleague, and it pulls what it needs from your Gmail and Docs. Two honest caveats before anyone gets excited. It needs the Google AI Ultra plan, which runs about $100 a month. And it is still in beta. The one piece of good news for this group: it is open to US users first, and your team is in the US, so you could get on the early list. I am in Europe and cannot yet.
Your homework, and a template to make it easy.
Pick one task you do over and over, and write a short SOP for it. Then try it in one tool you can get into today, Gemini in your Workspace or a free Claude or ChatGPT account. Not Workspace Studio or Spark, those may be locked for now. Bring what you tried, what worked, and what fell apart. It is completely fine if it does not work. Bringing the mess is the point. We fix it together next session.
One-page SOP template
- Trigger: what kicks it off.
- The steps: number them in plain order.
- What good looks like: how you know it worked.
- What it must never do: the hard lines.
Filled-in example
- Trigger: a new invoice lands in my inbox.
- Steps: 1 save it, 2 read the amount and date, 3 add a row to the expense sheet, 4 draft the accountant email.
- Good: every invoice on the sheet, email ready to review.
- Never: send the email or pay anything without me.