Welcome to the AI For Humans newsletter!

Nerdy tools have a pattern. They start out buried in a developer's terminal, confusing to those who don’t live in a code editor, and then quietly become the way everyone works.

Claude Code had this moment earlier this year.

Now, it's happening again with Codex, OpenAI's coding-first app that is very clearly positioning itself as something much bigger than a coding tool.

But there's a feature inside Codex right now that I think changes how AI works for you, and I'm not being dramatic.

It's called /goal.

I've been using it all week and it's vastly changed the scope of what I'm even trying to do with AI.

Let's get into it!

Please support AI For Humans by learning about our sponsors below:

Moda's viral launch hit 4.4 million views in two days. Tens of thousands of professionals signed up. Startups, agencies, forward-thinking brands and top firms are now using Moda to create brand-aligned slides, ad creative, reports, social carousels and more.

Most AI tools tend to create what we call "AI slop": repetitions of the same colors, layouts and fonts. And when you try to fix it, you get stuck in a loop of re-prompting.

Moda is different. Drop in your website URL, and Moda learns your brand from the ground up: your colors, your fonts, your visual language. Then it helps you generate pro-quality slides, docs, and marketing assets. 

The best part? Every layer is fully editable on a real canvas, and exports to powerpoint, PDF and more.

Try Moda Free Today

Ok so here's the quick version.

When you use Codex normally (or Claude Code, or Cursor, or any of these coding agents), you give it a prompt. It does a thing. It comes back and asks you what's next. You give it another prompt. Repeat.

That loop is fine for small tasks. Fix this bug. Write this function. Clean up this file.

But good lord, it is annoying.

But /goal is fundamentally different.

Instead of saying "do this one thing," you're saying "here's what I want to be true when you're done. Figure out how to get there."

And then it just... goes.

It plans its own sequence of steps. Executes them. Checks its own output. Corrects course when something fails. And keeps going until the goal is either met or it hits a wall it genuinely can't get past without your help.

If you've heard about the Ralph Wiggum loop (and if you haven't, you should look it up), /goal is basically the productized version of that idea.

/goal takes that same core idea (keep working, keep iterating, don't stop until the thing is done) and builds it directly into the product. No bash loop needed. No hacky workarounds. You just set a goal and let it run.

And when I say "let it run," I mean it.

I've had a project running now for over 24 hours straight.

I've steered it maybe five times total while it's been working. The rest of the time I've been doing other stuff, checking in on my phone, watching it chug along.

It's the first time using an AI agent has felt less like "prompting" and more like "managing."

And, that, dear reader, is what the AI CEOs have been promising forever.

Before I get too excited, I should flag something.

/goal running autonomously for hours means it's doing things on your machine (or in a cloud sandbox) for hours. And the permissions setup in Codex matters a lot here.

Codex has three permission profiles: read-only, workspace (writes only inside your project), and the ominously named "danger-full-access" which removes sandbox restrictions entirely.

If you're running /goal in workspace mode, you're prob fine. If you crank it up to full access, just know that security researchers have flagged real data exfiltration risks even with the default settings. The Codex Desktop App can read any file on your computer even in the most restrictive mode.

My advice: use the workspace profile, don't give it access to anything you wouldn't show a very enthusiastic but occasionally unpredictable intern, and check in regularly. The autonomy is the feature. The oversight is still your job.

Love this newsletter? Forward it to one curious friend. They can join in one click.

Ok here's where I want to zoom out a little.

If you're reading this newsletter and thinking "I don't code, why do I care about some command inside a developer tool," I get it. But bear with me.

What /goal is actually doing is giving you a preview of what all AI interaction is going to look like pretty soon.

Right now, most people use AI in a chat window. You type something, it responds, you type something else.

It's a conversation. And conversations are inherently short-horizon.

You're thinking one message at a time.

/goal is the opposite. It's long-horizon. You set an outcome and the AI works backward from it, planning, executing, testing, adjusting, for hours or even days.

It's the clearest early example of what everyone keeps calling "agentic AI," except it's not a demo or a concept video. It's a real thing you can use right now.

And here's the part that makes this moment different from even six months ago: GPT-5.5 is actually good enough to pull it off.

Don’t just take my word for it. Claire Vo made a good video you can check out below.

More powerful systems are coming. The models that land later this year are going to make this version of /goal look like training wheels.

But this is the starting gun.

We're moving from a world where you use AI to a world where you manage AI. You set the objective. You check in. You course-correct. The AI does the work.

And you (yes you!) should get comfortable with it.

-Gavin

This week on AI For Humans: New Opus 4.8 But Also…Mythos Incoming?👇

Jenny Krakovsky has bitten off an ambitious goal: make a full AI feature film in 30 days. And she's vlogging the entire process daily on X.

She's on day 8 and it's just cool to see someone go through this so transparently.

The wins, the struggles, the moments where the tools fight back.

If you were at AI on the Lot, you saw how many people are pursuing this. Jenny's the one actually showing her work, every day, in public.

Follow Jenny and follow along!