So you sit there trying to think of something, come up with nothing worth the effort, and go back to doing the same three jobs by hand you have done every week for a year. I did that for months. The list never arrived because I was trying to remember work I do on autopilot, which is precisely the work you cannot remember.

Then I stopped trying to think of it and asked ChatGPT instead. It has been watching me ask the same things over and over. It knows what I repeat better than I do.

It came back with eleven. Four were things I had genuinely done fifty times by hand.

The prompt is below. It takes two minutes and works on every plan, including the one that costs nothing.

The 20-second version: You cannot think of what to automate because the work you repeat is work you do without noticing. Your AI has a record of some of it. One prompt turns that record into a ranked list, and you can have the top item running by tomorrow morning. Works on every plan.

Time: two minutes for the list, five to set the first one running. You need: ChatGPT, with memory on, which it is by default. Payoff: something useful arrives tomorrow that you did not do.

The old way: you know you should automate something. You open a blank page. Nothing comes. The jobs that eat your week are invisible precisely because they are habits, and habits do not announce themselves.

The replacement: ask the thing that has been watching you work.

The line doing the work is number three. What stayed the same is the part a machine can do. What changed each time is the only thing you would still need to supply. Once you can see both halves, you know exactly what to hand over.

If the list comes back thin, memory is off or the account is new. Try again after a fortnight of normal use.

Take the top item, and use the instruction from point four. Open the Scheduled page in the left sidebar on chatgpt.com, paste it in plain language, and check the confirmation, especially the time zone.

If nothing jumps out and you have your calendar and email connected, start with this one:

Then turn on notifications under Settings, then Notifications. Without them, the task runs and nobody sees it, which is how most people decide this does not work.

That list is useful, and it is also incomplete in a specific way.

Your chat history only knows the work you brought to AI. The real record of your repeated work sits somewhere else: in your sent folder, your calendar, and your files. The email you have written forty times in slightly different words. The report you rebuild from last month's copy. The meeting you prepare for the same way every fortnight. You never mentioned most of it to ChatGPT, so it cannot see it.

And even the items it did find, most people then force into a scheduled message, because that is the only mechanism they know. This year added four more, and most of your list needs one of those instead.

Inside the Pro section:

The Automation Map. The first audit finds what you asked AI about. This finds everything else, and tells you how to automate each piece with what AI can actually do now:

Also in your library: the Connector Playbook, Google's AI Tools guide, the Skills Bank, and every workflow pack from past issues.

The team at Efficient App, who say they have tested more than 200 AI tools over three years, currently rank Viktor first. It works inside Slack or Teams, reads what is happening in your channels, and offers to take work off your plate before you ask. Their description is that it is the closest thing to an actual AI hire they have used.

It is the mirror image of this week's workflow. The audit has you asking the AI what you repeat. Viktor works the other way round: it watches what the team is doing and offers to pick things up.

Worth a trial if your work runs in Slack or Teams. Two honest caveats. This is one team's ranking rather than a broad consensus. And anything that reads every channel deserves the same care as any connector, so check exactly what it can see before you switch it on.

💸 One frontier model now costs 13 times more per token than another. A September comparison of GPT-6 Astra, Gemini 3.8 Flash and Claude Fable found Astra 13 times more expensive per token than Gemini 3.8 Flash (AI Agents Directory).

⏰ An older model retires on 14 October, and it will slowly break some automations. GPT-5.5 leaves ChatGPT, Work, and Codex on all plans that day, and OpenAI is telling users to review any scheduled task still using it and pick a replacement (OpenAI Learn).

🧩 A capable open model landed at a fraction of frontier pricing. Z.ai's GLM-5.3-Flash is the first natively multimodal model in its family, with a 1 million token context window, MIT-licensed weights, and list pricing of $0.15 and $0.50 per million tokens (LLM Stats).

🖥️ Local AI hardware is arriving in October. NVIDIA says RTX Spark Windows PCs ship this month, built around local inference, on-device agents and running models without a cloud connection (CodeMicros).

Quick version: A September comparison found GPT-6 Astra costs 13 times more per token than Gemini 3.8 Flash. Four labs shipped in one week and buyers say the pace is exhausting, so almost nobody has revisited a choice they made a year ago. If you run anything repeatedly, you are likely paying several times more than the job needs.

Most people picked a model once. It was the best one at the time, it became a habit, and every task since has gone through it. Summarising a document, drafting an email, extracting figures from a PDF, formatting a list, all running through the most expensive option available, at the slowest speed, often for no measurable gain.

That was a reasonable position while the models were converging. It stopped being reasonable when the price gap opened to 13 times (AI Agents Directory).

To be clear about what that gap is not, Astra is genuinely better at some things. It is the strongest model available for operating software, multi-step work and anything spatial or mechanical. If that is your job, pay for it.

But that is not most of what most people do. Most of what most people do is read this, summarise that, draft the other. On those jobs the cheap models have got good enough that a blind comparison is uncomfortable, which is the test in this week's workflow and worth running.

The thing that makes this urgent rather than merely interesting is the compounding. A daily scheduled task on the expensive model is 365 runs a year of something a cheaper one would have done identically. That is where a preference quietly becomes a line item.

And the pace is not helping anyone. Four frontier releases in one week, with buyers openly describing the tempo as exhausting. The honest response to that is not to keep up. It is to stop choosing by brand and start choosing by job.

🔒 The Pro breakdown: the independent benchmark result that makes the expensive models hard to justify, the full price ladder, the answer for every common job, and what it costs you in real money over a year.

Workflow Library Updates:

NEW Resources:

Each resource lives permanently in your Pro account. Use them whenever you need them.

The bit that stayed with me was not the automations. It was seeing eleven things written down that I had never noticed I was doing.

You cannot plan around work you do on autopilot, and you cannot remember it either. But it has been recorded, so ask for the list.

Two minutes for the audit, five to set one running, and tomorrow morning something arrives that you did not do.

👇 What’s behind the paywall:

Pro members deploy AI in their work an average of 5-8x more often than free readers (based on reply data from past issues). The difference is having the exact setup, not the concept.