Dead Pixel Design
2026-10-18

Ambition Has a Budget

An Anthropic engineer wrote a guide to being more ambitious with AI. Here's how it plays out in a one-person web studio with a usage cap and real clients.

aiagentsworksmall-business

Earlier this year Codex built me a little kanban board that kicked off a headless orchestrator agent. I gave it a lead-scraping job and let it run. It burned through a full five-hour Claude Max usage window in under six minutes.

I took it apart.

I still like the idea. I just learned something about ambition that doesn’t show up in most of the posts telling us to have more of it.

Last week Felix Rieseberg, an engineering manager at Anthropic who’s led engineering on Claude.ai and Claude Code Desktop, published How to be more ambitious: A tutorial for humans with models. It’s good, go read it. He breaks ambition into four kinds. Speed, doing the same work in less time. Altitude, handing the model the outcome instead of the task. Depth, working the problem the way the best person in the field would. Range, pulling in fields the task wouldn’t normally touch.

I agree with almost all of it. But he’s writing from inside the building where the models get made. I’m one guy running a web studio on a couple of subscriptions, building sites for small businesses and nonprofits that don’t have an IT guy. Ambition looks different from here. It has a budget, and there’s a person on the other end who has to live with whatever I build.

Speed is the only one anybody’s selling

Felix skips speed because it’s the boring one. Fair. It’s also the only one small business owners are getting pitched. Website in a day. AI site builder, launch in minutes.

None of that was ever the problem. A site built in an afternoon still has no real content on it, still says nothing a customer in the next town would care about, still isn’t hooked up to anything. Faster just gets you to the same empty page sooner.

So I mostly ignore speed too. Not because it’s boring, because it’s not where the work is.

Altitude starts with plumbing

The best line in Felix’s post isn’t the ambitious one, it’s the one right after. He stopped asking Claude to fix individual crashes and started asking it to keep the apps crash-free, but first he had to give it the crash reports, map app versions to commits, and hand it actual Windows, Mac and Linux machines to reproduce and verify fixes on.

Altitude is mostly plumbing.

I learned that with REAPER. I wanted a model to actually work a mix in my DAW, not just talk about one. So before anything else I built Reaper Daemon, a local JSON file bridge that lets an agent read and control a REAPER session with no network layer in the middle. Once that existed, GPT-6 Astra mixed one of my songs in REAPER and handed the mix back in an audio player. The mix was the easy part. The bridge was the work.

Client work is the same move at a different altitude. A nonprofit in a small Maine town hired me to build a business directory for their area. The task was “a directory website.” The outcome they actually needed was for the town to keep that directory current without calling me every time a business changes its phone number. So it got an admin dashboard where they edit listings, approve events people submit, and approve businesses claiming their own page. The site was the task. Them running it without me was the goal beyond the goal.

Altitude without a stop button

Back to the board that ate my usage window.

That was altitude. I gave a model an outcome and let it work out the tasks. It worked out a lot of tasks, very quickly, and spent everything I had doing them.

Felix mentions there are limits (“Claude, go make money” doesn’t work for most businesses yet). The limit I hit wasn’t intelligence, it was supervision. Inside a company that size a runaway agent is a line item. In a solo studio it’s your whole day’s capacity, gone before you finish reading the log.

So the version I’m putting together now inside my own desktop app starts from the other end. The orchestrator watches token spend while it works and corrects before it gets out of hand. There’s a kill switch that actually kills. I can talk to it mid-run and change direction. No swarm of the most expensive model doing what a cheaper one could handle.

Raise the altitude, sure. Build the brakes first.

Depth means looking at the raw stuff, including your own

Two of Felix’s depth techniques hit home. Challenge what you know, by giving the model the raw material and asking it to check your assumptions. And use an outside judge, because a model grading its own work will usually give itself an A.

I pointed both at my own studio. A brand audit of Dead Pixel Design’s assets turned up that the brand guide in my Drive (Playfair and Inter) and the live site (Cormorant and IBM Plex Mono) were describing two different companies. I’d have told you they matched. And for the site redesign I wrote an adversarial front-end reviewer prompt for a different model, Kimi K3, whose whole job is to tear into the design instead of agreeing with it.

None of that is expensive. Depth on a budget is mostly being willing to be wrong about your own stuff.

Range is where a small shop is already strong

Felix borrows a trick from The Art of Innovation by Tom Kelley and Jonathan Littman, replacing the noun with a verb. A quarterly review isn’t a deck, it’s getting busy people to believe a story about cause and effect and decide in 30 minutes.

Do that to a small business website. It isn’t a website. It’s getting somebody two towns over who’s never heard of you to trust you enough to pick up the phone. Once it’s a verb the crew changes. You need a copywriter more than a developer, somebody who thinks about local search, somebody who’ll read it like a skeptic.

A one-person shop has range because it has to. I write and record metal, produce in REAPER, and build websites, and the REAPER bridge came out of the same toolbox as the client sites. What the models change is that I can actually fill the other seats now, as long as I ask the model to do what that person does instead of talking like them. Felix makes that point too and he’s right. Role play makes the transcript weirder and the work no better.

The ceiling is the person you hand it to

Felix closes on self-efficacy, the idea that we’re held back by what we believe is possible, and that watching someone else do the impossible thing is one of the ways past it. I buy it. Watching a model mix my own song changed what I think I can build.

But the people I build for aren’t chasing a four-minute mile. They’re running small businesses and nonprofits with no IT department and no spare hours. The most ambitious thing I can build for them isn’t the most impressive thing a model can produce. It’s the thing they can still run a year after I’m done.

Ambition that only works while I’m babysitting it is just a really good demo.

So I’d add one thing to Felix’s list. Not a fifth direction, more a test you run on the other four. Does it survive the handoff?

He ends by asking people to send him more techniques. Mine is short. Build the stop button first, and build the whole thing for whoever’s left holding it.

Related

Keep reading.