AI with Kyle Daily Update 191: Don't Automate (Yet)

Greg Brockman (co-founder of OpenAI) shared a self-improving Codex prompt:

It’s a bit misleading if you stop at the Codex bit.

This is bigger than Codex. This is useful regardless of what AI you are using. And indeed we can go wider and do even more with this.

The prompt asks your AI to look back over your recent work, find repeated manual workflows, check existing skills / agents / automations, and then create the smallest useful missing piece. A skill. A subagent. An automation. Or nothing, if the evidence is a bit crap.

Here’s the full prompt btw:

I ran this on my livesteam and it immediately spotted obvious repeat loops in my own work: slide decks for livestreams, turning messy source material into public explanations, AI tells checks, system diagrams, short-form explainers, repurposing.

BUT We can do more. We can use AI to find work you keep doing by hand.

We’re not talking about your fantasy productivity system. Not the make-money-whilst -you-sleep agent setup Twitter wants you to build. Your actual week. The boring stuff that you do on the day-to-day. The things you repeat, delay, copy-paste, explain again, or quietly dread every month.

That is where we can actually make our lives better.

Start With YOUR Real Work

Do not automate for the sake of it. Good lord.

First work out what you actual should be using AI on.

The first move is an audit.

Not a fancy one. You do not need a 14-tab productivity dashboard called LifeOS (please don't). You don’t need to vibe code up a SaaS just to get this done.

You need to look at the last 7 to 30 days and ask where your time actually went.

Calendar. Inbox. Docs. Notes. AI chats. Browser history. Timesheets if you have them. Pen and paper if you must. Absolutely nowt wrong with crude capture if it gets the job done. Scribble it down on the back of an envelope - does not matter.

The point is to look at your real week, not your ideal week.

The prompt above is great but has a weakness. The AI only knows what you have already shown it. If all your repeated work lives in email, WhatsApp, spreadsheets, meetings, client calls, or the weird five-step admin thing you do on Friday afternoons, ChatGPT will not magically see it unless you bring that evidence in. You’ll be shit out of luck for all the stuff outside of AI - stuff that you probably should be using AI for.

If you ask AI "what should I automate?" too early and you get a narrow answer.

It will answer from its own little window. It’s working in its known knowns - it has unknown unknowns that it can’t account for. And that window might miss half your life.

The Best Candidates Are Usually Boring

Start with boring stuff.

My example on stream was bookkeeping.

Once a month, Harms sends me a sheet. I look at it and think, "Oh no, this again." It takes too long. It blocks other people. I do it repeatedly (and badly). The inputs are similar. The output is similar. The steps are similar enough.

Perfect candidate.

Not exciting. Not viral. Not a cool "AI agent that runs my business while I sleep" thing.

Just annoying.

And annoying is useful.

Client updates. Meeting notes. Call summaries. Proposals. Analytics checks. Competitor research. Common replies. Report formatting. Turning ideas into briefs. Taking messy notes and making them into actions.

That is the good stuff.

Because if you do it more than once, and you will do it again, even saving ten minutes each time can be worth it. Especially if the task is error-prone, context-heavy, or causes a bunch of faff around the edges. And those little wins gain momentum as you stack them up.

Think of Atomic Habits but with AI agents…wow, that’s a bestseller idea right there. Huh.

Use AI As The Interviewer

If your work is already inside AI that viral prompt can go hunting through the history.

Lovely.

But if your work is spread across your actual life one more tip to capture everything is to put the AI into interview mode.

This is one of my favourite ways to use AI because it removes the self-editing. You do not need to sit there like a Victorian clerk writing a perfect process document. You can just talk.

Use something like:

Ask me questions about my last 7 to 30 days of work.

Help me spot things I repeated, delayed, copied and pasted, or explained more than once.

Then help me decide what is worth turning into a reusable prompt, checklist, template, automation, or small tool.

Then answer verbally if you can. You’ll be less likely to “catch” yourself. Just blab. Verbal diarrhoea. (Wow that took me far too many attempts to spell correctly, even with spellcheck…I wasn’t even close…)

The microphone is underrated for this. You get more out of your head because you are not trying to make the sentence nice before it exists.

Pick The Smallest Fix That Works

Stop right there. I see you. Reaching for a complicated (but oh so cool) solution.

People overcomplicate the fix. Especially us AI nerds.

You do not need an agent for every repeated task. Sometimes the answer is a reusable prompt in a doc. Sometimes it is a checklist. Sometimes it is a template. Sometimes it is an automation in Make or Zapier or n8n. Sometimes it is a small script.

Sometimes, yes, it might be a managed agent, Claude Code subagent, Codex setup, OpenClaw thing, Hermes thing, whatever.

Even if it is…start small.

If the repeated thing is the same kind of answer, make a prompt.

If it is the same steps each time, make a checklist.

If the shape of the output matters, make a template.

If several tools need to talk to each other, then maybe build an automation.

If it is a specialist recurring investigation task, maybe create an agent or subagent.

If it is a fiddly internal process that does not need AI every time, write a script or tiny app.

That's it.

No need to build a self-improving cloud creature because you send the same client update every Thursday. That’s overkill. Use the lightest thing that makes next time easier. If it proves useful, improve it later.

Keep Asking The Small Question

This whole thing collapses down to one boring question:

What did I do this week that I will probably do again?

That is kinda it. But not enough people actually bother with this introspection.

Ask that every Friday. Or every Monday. Or whenever you are already faffing with planning and pretending it is work.

Write down the repeats. Pick one. Make next time easier.

Maybe that means a prompt. Maybe a checklist. Maybe a template. Maybe a tiny automation. Maybe nothing because the task is rare, messy, or too personal.

Fine.

Skipping is part of the system. Otherwise you end up building a stupidly elaborate automation for a task you do twice a year, which is basically procrastination wearing a tiny developer hat. Feels productive but it’s the opposite.

To the Task,

Kyle