Sort incoming topic ideas

So far you have briefed AI one task at a time: you write a prompt, get an answer and check it. Mark Zuckerberg, Meta's CEO, describes the shift this course is about. With a chatbot you send one prompt and get one answer. With agents, "you give it goals", and they keep working until the goal is met.

A proactive agent starts the work itself, on a schedule (every Monday) or when something happens (a new export lands), without waiting for you to prompt it. That is useful, and it is also why the goal needs clear limits.

In this course you write each goal as a standing goal card with six parts:

1. Role name, for example Editorial Desk – Intake
2. Scope: what it does and what it leaves alone
3. Approved sources and connections
4. The agreed report: what it hands you, and in what format
5. The human owner: the person it answers to
6. Stop, pause and retire rules

The Monday prompt "sort these comments" gets done once. The card Editorial Desk – Intake can run every week, and you know where it stops.

Before you build any agent, decide which jobs it should do. Jim Louderback, former CEO of VidCon, suggests a simple test: go through your last twenty posts and ask, for each one, whether an AI could have made it. His advice is to "lean into the no".

Do the audit on the preparation behind each post as well as the finished post. For the cliff-walk answer, the work broke down like this:

• Yes pile (an agent can prepare it, and you check it): sorting the week's questions, building research packets, first script drafts, caption files.
• No pile (stays yours): the walk you took in May, your nephew sitting down on the 200th step, the word you give your audience, and the last look before anything goes out.

The yes pile becomes your list of standing agent jobs. This lesson builds the first one, intake.

Louderback also suggests a digital twin for the rest. This course replaces that with a firm rule: no agent ever uses your face or voice. Agents prepare the material. You stay the person on screen and the author.

Here is a filled-in card for the first job from your yes pile.

• Role name: Editorial Desk – Intake
• Scope: Read the audience questions in the approved exports, group them by viewer need, and write the report. Don't reply to anyone, don't rank by reach and don't research the answers.
• Approved sources and connections: This week's comment exports from TikTok, Instagram, YouTube, Snapchat and the Facebook group, saved to one shared folder, plus the submitted idea inbox. Nothing else.
• Agreed report: A question-clusters PDF, delivered to the owner.
• Human owner: You. If you have a part-time editor, still write down one name. On a one-person channel every role is yours, and it is still written down.
• Stop, pause and retire: Stop and report if an export is missing or contains private messages. Pause when the owner says pause or when the series takes a break. Retire when the series ends, then disconnect the folder.

Steven Bartlett starts every hire by asking whether an agent could do the job. This course adds the reverse question: which person does each agent answer to?

You make two choices.

The tool. OpenAI's Dots and Anthropic's Claude Duet are taught here as equals. Each runs on its own cloud computer, connects to the apps you approve, works toward a goal while you are busy, and checks with you before sending, paying, signing or deleting. Choose by where your exports actually live. If your comment exports go to one drive or inbox, pick the tool you can connect to it with the permissions you're comfortable giving. Open the tool's current settings and check what it can connect to and when it asks you. Don't assume a feature exists because another tool has it.

The run type.

• A single run does one pass, such as this week's comment export, and ends with a report.
• A continuing brief runs on a schedule, such as every Monday morning. Because it keeps going without you, it needs written pause and retire rules from day one.

In practice, test the role as a single run first. Once a report has passed your line-by-line check, turn the same card into a continuing brief.

Write one sentence giving your reason for each choice. You will need those reasons when you review the agent later.

An intake agent only needs the questions. Connect it to approved exports of audience questions, such as comment exports saved to a folder, or a submitted idea inbox where viewers send questions on purpose. Never give it full access to your DMs. Private messages, and anything about children, should stay out of cloud tools unless the answer truly needs them.

The cliff-walk family has children aged six and eight, so this is a real concern. Write it into the scope: keep the need, drop the identity.

Original comment (from a named account, with a profile photo):

"Can my two (6 and 8) manage the cliff walk next month? And how do we get back to the car?"

What the agent keeps: children aged six and eight; 300 steps near the end; getting back to the car; visiting next month.

What it drops: names, handles, home locations, photos and any private-message thread.

That is enough to plan an answer and too little to identify the family. If an export arrives containing DMs, the stop rule applies: the agent halts and tells you, and doesn't sort around them.

A viewer need is the real question under the words. The parent literally asked "can they manage it?" Underneath are two needs:

1. Will the 300 steps be too much for young children?
2. Will we make it back to the car before the last shuttle?

The agent labels each cluster with a need like these, written as the viewer's question. It doesn't use a topic word ("cliff walk") or the loudest phrasing from the comments.

Two rules go in the scope:

• Never rank by provocative wording or predicted reach. A comment shouting "DEATH TRAP for kids??" belongs in the same cluster as the calm parent, and it doesn't move to the top for being dramatic. An agent that guesses which clusters will "perform" will drift toward outrage.
• The order of the queue is your decision. The agent lists the clusters. You decide which question to answer first, based on what your audience needs and what you can honestly answer.

A line for the card: "Label clusters by the viewer's underlying question. Do not sort or score by emotional wording, engagement or predicted reach. List clusters in no ranked order; the owner sets priority."

The agent's output is a question-clusters PDF. For each cluster it shows:

• Viewer need, written as the viewer's question
• Number of questions in the cluster
• Platforms they came from
• Paraphrased sample questions, with no names or quotes that identify anyone
• Open questions for you, such as "Two comments mention dogs; out of scope or a new cluster?"

Test it before you trust it. Run the card as a single run on a small export, for example last week's comments from two platforms. Then check line by line: open the export and the report side by side and confirm every question landed in the right cluster, none were lost, and nothing identifying slipped through. If the steps question and the shuttle question got merged into one "cliff walk" cluster, the labels are topics, not needs. Tighten that instruction and rerun.

Only when a report passes this check should the card become a continuing brief. The next lesson deals with the questions that come back again and again, and with ideas that have gone stale.