Good morning.

Last month I asked what one job you'd hand to AI tomorrow. I got a lot of replies.

Nearly all of those jobs run on the Claude or ChatGPT plan you already pay for, and the ones that go wrong tend to go wrong at the same moment, when Claude or ChatGPT sends, posts or publishes before you've looked.

In today’s issue: every job, what to use for it, the smallest version you can run this week, and what it costs to set up.

—Sam

A lot of you filled in the September survey, or answered the poll in How Your Agents Give You an Unfair Advantage.

Nobody picked "I haven't really started yet." Across both, almost all of you use Claude or ChatGPT for one-off tasks, some have built a few repeatable prompts, Skills or workflows, and many run agents or automations that work without you.

Asked to pick one area, the most common answer was agents and automations that run without me. That matches the jobs below. Most of them are something you do every day or every week that you'd like to stop doing by hand.

Each answer below has four parts:

If you're a coach or consultant working on your own and want one place to start, start with the FAQ drafts or the expense report. Both pay back in the first week and neither needs anything connected.

"I've got work plans for multiple projects in Teams and I want to generate an email to every team member letting them know what they have to accomplish every day."

The plan already holds everything each person needs to hear. The work is reading it from each person's side every morning, which is exactly the kind of reading Claude or ChatGPT does well and a person does badly at 7am.

Start with an export. Whether your plans live in Planner, a Teams-hosted spreadsheet or a list, get them into one sheet with a row per task: project, task, owner, due date, status, and what it's waiting on.

Paste that in and run this:

The WAITING ON YOU section is the reason to bother. A list of your own tasks is something people already have, and what they don't have is the view of who's stuck behind them.

Once it runs clean for a week, make it a scheduled task. In Claude Cowork or ChatGPT Work, with the Microsoft 365 connector turned on, point the scheduled task at the plan file and have it leave one draft per person in Outlook.

Keep it at drafts. I'd want a fortnight of drafts that needed no edits before anything sends on a timer. When you get there, Microsoft's own Power Automate can do the sending, and Claude or ChatGPT keeps writing the text.

If the export is the step you keep skipping, hand it to Claude Code or Codex. Ask for a short script that pulls every project's plan into the one sheet each morning, so the scheduled task always starts from a fresh copy. Either one will walk you through the Microsoft 365 access the script needs, which you grant once.

Cost: an hour to set up the export and the prompt. Nothing beyond your plan.

"Building my expense reports."

Short answer, common job. An expense report is a pile of receipts sorted into a table your accountant already defined, which makes it close to the PDF-to-spreadsheet prompt from Eleven Jobs. This version adds the parts that are particular to expenses: categories, duplicates, and the purchases that might be personal.

Drop the month's receipts into one folder, as photos, PDFs or forwarded email receipts. Claude and ChatGPT both read a phone photo of a crumpled receipt well, so you can snap them as you go. Then upload the folder to either one and run:

The ASK ME flag is what makes it safe to hand over. Left alone, Claude or ChatGPT will categorise a family dinner as meals with clients with full confidence, and that's the line that causes trouble later.

After two months of clean runs, save it as a Skill called expense-report, or as a ChatGPT project with the instructions in it. The job then becomes "run my expense report for October" and a folder. If your accounting software takes a CSV import, match its columns and the job ends in an upload.

Cost: an afternoon. Nothing beyond your plan.

"Responding to FAQ emails."

This came from a consultant who already runs agents, so I'll skip the basics. I printed a support-reply prompt in July, in Get the Most Out of Fable 5. The newer, better part is building the answers file from your sent mail instead of writing it from scratch, then having the drafts wait in your inbox.

Step one, build the file. Run this once against the last few months of your sent mail:

Read the file, fix the conflicts, delete anything you'd answer differently now. That's an hour, and you end up with your own answers in your own words.

Step two is the scheduled task, in Claude through its Gmail connector or in ChatGPT through its Gmail or Outlook connector. Have it read new mail every morning, match each question against faq.md, and write a reply using only the file.

If your connector can create drafts, leave each one as a draft in its thread. If it can only read, have it produce every reply in one document under the email's subject line, and paste them yourself. Anything the file doesn't cover gets a label, NEEDS ME, and no draft. Nothing sends.

"Honestly, getting through the junk in my email and the real stuff in my email."

This one's from a freelancer and it's the same job. The daily desk prompt in Eleven Jobs does the sorting. What it lacks is your definition of junk, so add a short file the prompt reads first:

Label instead of deleting for the first month. You'll find a couple of wrong calls, and you want them in a folder you can check.

Cost: two hours for the FAQ file and the scheduled task, thirty minutes for the junk rules. Nothing beyond your plan.

Several of you asked about content:

"Schedule events over multiple social media platforms."

"Automate my daily blog posts with images and links."

"Good content writing & workflow."

"Take a creative brief and produce a 9/10 quality end product without me touching it."

In Eleven Jobs I said I wouldn't write a fourth issue on content writing unless something moved. Something did. Claude and ChatGPT can now reach the scheduler and the blog platform directly, through connectors, on a timer. So the useful question now is about plumbing: where the source material lives, where the drafts go, and where you look before anything goes public.

Keep one sheet with a row per thing you're promoting: date, title, link, image file, the platforms it goes to, and one line on why someone should care. Then run this in Claude or ChatGPT:

That CSV is the hand-off. Most schedulers take a bulk CSV upload or have their own connector, so check which yours has. I queue mine to Typefully through its connector and everything goes in as a draft I approve. Either way, the posts sit in the scheduler waiting for you. That's the step to keep.

The plumbing here is a scheduled task that reads a topic list, writes the post, adds images and links, and leaves it as a draft in WordPress, beehiiv or whatever you publish on. Two files keep it honest:

For images, use your own library first. When you need something new, ChatGPT's image generation and Gemini's image model both handle a clean header image from a one-line description, and Midjourney is worth it if you want one consistent look across every post. Have every image come with alt text.

For the drafting and loading, Claude reaches WordPress and beehiiv through connectors (it's how this newsletter gets loaded into beehiiv), and ChatGPT can reach the same platforms through a custom connector in Developer Mode. Then publish by hand, because pressing publish is the one moment you see the post the way a reader will.

Two clients' results answer this better than I can.

A freelance copywriter I've worked with built one Claude Project per client with twelve to eighteen of that client's best pieces in it. Across sixty-one deliverables, revision rounds fell from 2.4 to 1.1 on average.

She'd spent three weeks describing each client's voice in prose first, and none of it produced anything usable, while twelve real examples worked straight away. Her time savings are self-tracked, and she admits two clients' "before" figures were estimated from memory.

A freelance video editor turns each raw transcript into an edit plan and sends the plan for approval before he edits anything. Across forty-seven videos, revision rounds went from 2.8 to 0.9. The catch, which two of his clients have mentioned, is that they now do their reviewing earlier.

Without me touching it is the wrong target. Touch it once, early, at the plan, and give it samples rather than adjectives. This is the plan-first prompt:

Cost: an hour for the social sheet, a day for the blog pipeline, an afternoon to collect samples. Your scheduler is the only extra.

"I spend a lot of time creating and organizing sales pages, ad creatives, copy, prompts, product content, and funnel assets across different projects. A system that could take a raw idea or research and automatically turn it into a structured, ready-to-launch digital product + marketing funnel would be incredibly useful."

You can build this, and the order decides whether it works. Most attempts generate the sales page, the ads and the emails side by side from the raw idea, so each asset invents its own version of the product, and you spend your time reconciling the sales page against the ads.

Put one file in the middle. Everything gets written from the brief, and nothing gets written before you've approved it.

Gather the research with a tool built for it. Perplexity is the fastest way to get market and competitor material with a source on every claim, and ChatGPT's deep research or Claude's Research mode will give you a longer report from the same question.

Add what you already have, which is usually better than anything they find: customer emails, call notes, survey answers, the questions people ask in your DMs. Then paste it all into a Claude Project or ChatGPT project and run:

Once you've approved the brief, every asset gets its own prompt that reads the brief and nothing else: sales page, five ad angles, the launch emails, the product outline. Add one rule to all of them: every claim about what the buyer gets must match DAY ONE. If it doesn't, cut it.

A consultant I work with used that split to presell a $1,490 group program that was about 40% built. Her sales page listed what was available on day one, then separately listed what would be built during the cohort, with dates, and her emails could only read the first list.

She sold thirty-eight seats against a target of thirty, for $56,620, with refunds at 7.9% against her usual 11%. It's one cohort, and her previous cohort ran at a different price, so the comparison isn't clean. Nine testimonials mentioned the honesty about what was built, which she never asked about.

Set it up as one Claude Project or ChatGPT project per product, with brief.md at the top and every asset prompt saved inside, so the next product starts from a copy of the folder. If you want the version that tests the offer before you build it, that's what the Offer Kit in Cortex was built for.

Cost: a day for the first product, mostly spent reading the brief properly. Nothing beyond your plan.

A few of you asked about ad data, and one of you described the problem exactly:

"I got ChatGPT to set up Looker Studio for me but I was not able to blend the data from Meta with GA4, so I still have to ask it questions to know which ad has the lowest cost."

The blend fails because Meta and GA4 don't share an ID. Meta records which ad was clicked and GA4 records what the visitor did next, and nothing connects the two unless the link passes the ad's identity into GA4. Meta can do that for you. In each ad's tracking settings, click Build a URL parameter, set Campaign content (utm_content) to Meta's dynamic ad.id value and Campaign name (utm_campaign) to campaign.name, and Meta fills them in on every click. From then on, both exports share a column.

Then skip the blend. Export Meta's ad-level report and GA4's report by utm_content once a week, upload both CSVs to ChatGPT or Claude (both run the numbers in code, so the arithmetic holds), and run the two-system reconcile prompt from Eleven Jobs, joining on the ad ID. Ask for cost per ticket link click per ad. Check the match rate first, since ads from before you added the parameters won't match, and that's expected.

For deciding what to cut and scale, a coaching business I work with, spending $31k a month, has an agent that sends one page every morning: cut, scale or leave for each ad set, with 3-day, 7-day and 14-day returns side by side. It has no access to the budget, and the founder decides.

Her blended cost per customer fell from $412 to $317 over five months while spend rose 19%, though she also changed a landing page in July and doesn't credit the agent alone. The smallest version is a weekly prompt you run in ChatGPT or Claude against the export, and once it's useful, a scheduled task that runs it every Monday:

The fuller ads readout, with the business case built from campaign and market data, is on my list for early next year.

Cost: two hours to set the parameters and the first week's prompt. Nothing beyond your plan.

Some of you do work specific enough that no general prompt will do it well: estimates for design-build remodels, uploading finished products to Amazon, and technical reports. All three run on the same pattern:

For remodel estimates, the rules file is your cost book: unit costs, labour rates, markup, contingency, and the line items you always forget. Then:

For Amazon uploads, the rules file matters even more, because the marketplace enforces it. A home goods seller I work with had an agent writing listings and buyer replies with no platform rules in it. Nine of 612 replies told buyers to email the business directly, which Amazon prohibits, and fourteen of thirty-eight rewritten titles ran over the category's character limit and got cut off in search. The account got a policy warning.

The fix they're putting in is a platform rules file (banned phrases, character limits per category) and a check that fails anything breaking it. Build yours that way from the start. For the upload itself, have Claude or ChatGPT fill in Seller Central's inventory file template for your category from a product sheet, run the check, then upload the file yourself.

For technical reports, the long-files prompt in Eleven Jobs does the drafting, with every claim cited and anything inferred marked. Keep it in a Claude Project or ChatGPT project with your report template, and add two past reports you're proud of as samples, which is the copywriter's lesson from the content section again, and the drafts start sounding like your reports.

Cost: an afternoon each, most of it spent writing the rules file. Nothing beyond your plan.

These two came with one line each, so I'll answer with what I have and tell you what would make the answer better.

"Make more sales for my digital product subscription."

On a subscription, a subscriber who stays is worth more than a new one who leaves in month two, so start by finding out which of your traffic sources bring buyers who stay. Export your subscribers with signup source, start date, cancel date and cancel reason, upload the file to ChatGPT or Claude, and run:

Spend more where the stayers come from, and fix whatever the first-month cancellers are telling you. Reply with your price, your monthly signups and your churn rate, and I can tell you which of the two will move revenue faster.

"Handle traffic (ads, influencers, press and so on)."

The tools split the prep between them, and you keep the relationships:

For influencers and press, the useful prompt researches before it writes:

Ads are covered in the section above.

One caution from a firm I work with:

They automated the front of their funnel, and qualified opportunities went from nine a month to twenty-one while their close rate fell from 44% to 23%. They put it down to delivery falling behind, with new clients waiting weeks to start, though a bigger funnel bringing in weaker leads probably explains part of it.

More traffic only helps if the next step can take it, so check that step before you turn the volume up. Reply with where your traffic comes from today and what happens after the click, and I can be more specific.

A few of your jobs have a whole issue behind them already, so here's where to find them.

One tap tells me which kind of business to build the examples around in the next one of these. If yours doesn't fit a box, pick Other and hit reply to tell me what it is.

Most of what you sent me is work you do every day, on your own, that nobody pays you for directly:

The morning sort, the receipts, the same answer to the same question, the post about the thing you already announced.

That's the time you get back first, and on a business your size it's the time that goes into selling and delivery.

Find your job in the table at the top and run the smallest version this week. If yours isn't here, or the answer missed your situation, hit reply with the job and a sentence about your business. That's where the next one of these comes from.

In Cortex, I share the formulas and strategies that help busy online business owners implement AI and Agents inside their businesses, in just minutes or hours instead of days or weeks.

Learn more about Cortex here

Talk soon,
Sam Woods
The Editor