What's Possible Beyond ChatGPT

What's Possible Beyond ChatGPT

Take one thing you do every week and walk it from chat, to co-work, to code – with the honest ceiling of each level.

You Don’t Climb
By Learning More

Most people I talk to have used ChatGPT for a year. They are good at it. They run a business, or a function inside one, and they have simply never had a reason to go past the chat window – nobody has shown them what is on the other side of it.

When they ask me where to start, they expect me to name a tool. Sometimes a course. It is neither. It is a question about their own week.

Because the standard advice – read the newsletter, watch the tutorial, try the tool everyone is posting about – produces almost nothing. A skill you learn with no task attached has nowhere to land. You close the tab and your week is identical.

You climb by handing over one task at a time.

A task beats a skill for three reasons, and they compound. It carries its own context – the tools it lives in, the people who receive it, the data it needs – so handing it over forces you to connect AI to your real work, which is the whole distance between Level 1 and Level 2. It has an obvious test: either it got done or it did not. And it recurs, so the payoff arrives again next week without you doing anything a second time.

Which is why one task at three levels teaches you more than three tools at one level. You already know what good looks like for your own work. That is the expertise this whole thing runs on, and you have had it the entire time.

If what you want is the full map rather than the method, I drew that separately. The AI Skill Path is six levels, two tracks and forty-eight capabilities, and it answers where am I. This is the three-level cut, and it answers how do I move.

Chat, Co-Work,
and Code

Each level gets the same four questions: what it is, what it unlocks, where it runs out, and what the next one fixes. The third is the one that matters. Without an honest ceiling, levels are a list. With one, they pull.

Chat – Your thinking partner
Level 1

Chat

– Your thinking partner

A better Google. It researches for you, drafts for you, and thinks with you – from what it already knows.

This is where almost everyone is, and it is not a bad place to be. A chat window is the best thinking partner most people have ever had at their desk. It is not weak. It is blind.

What it unlocks

Better thinking, faster. More personal answers than any search engine gives you.

Where it runs out

It has never seen your actual work – not your inbox, your calendar, your documents, your customers.

What the next level fixes

The next level gives it access to the work itself.

Before you go looking for the next level, take everything this one has. This is the single change that gets the most out of chat – and it makes the ceiling easier to feel, because you will have stopped blaming your prompt for it.

I want to use you as a thinking partner for my business, not as a search engine.

Before you answer anything, ask me the five questions you most need answered to be genuinely useful to me – about what I actually do, who I sell to, and how the money works.

Then stop and wait for my answers. Don't guess any of them, and don't start advising until you have them.

Are you at the ceiling already?

Tick the ones that are true of your last month. Nobody sees this but you.

Now pick yours

Reading about three levels is the easy part. Naming the one thing you will actually hand over is where it gets real, and it takes about a minute.

This is the same walk the live session does together – the room votes on a task and I run it up all three levels on screen. Here you choose your own. Do it now, before you read the next two levels: they land differently when you have a specific job in mind.

01

Pick one thing you do every week

It should recur at least weekly, and you should quietly resent it.

These shape the shortlist for the next session. Yours might be the one we build live.

02

Which is closest to your situation?

03

Walk it up

Pick a task and a situation and the three levels open up, one at a time.

Co-Work – Your colleague
Level 2

Co-Work

– Your colleague

Connected to the suite you already work in – mail and calendar, Teams or Workspace, Linear or Asana, Notion, your analytics, and the browser itself.

Three words do the work at this level, and all three get thrown around without ever being explained. A connector is a permission slip: you grant AI access to one specific place your work lives – your calendar, your inbox, a folder, a project tracker. A skill is a written-down way of doing one job, so the good version becomes the default instead of something you re-explain every week. An agent is simply the assistant once it can act rather than only answer. None of this involves code, and none of it produces code. It produces finished work.

What it unlocks

Your better thinking turns into finished work. Requests get turned around faster and come back at a higher standard.

Where it runs out

An agent is automating the work, but it still waits to be asked – every time, from scratch.

What the next level fixes

The next level stops automating the task and starts building the tool that does it.

Level 2 automates a task. Level 3 manufactures an instrument.

The prompt that starts the move. Notice it asks about your stack instead of assuming it – one that presumes Notion and Linear is useless to someone on SharePoint and Jira, and reads as though we never thought about them.

I want to hand you one recurring task instead of doing it myself every week.

The task is: [describe it in one sentence].

Before proposing anything, ask me which tools it touches – where the information lives, where the output has to end up, and who else sees it. I'll tell you what I actually use; don't assume any particular product.

Then tell me the smallest set of connections that would let you do it end to end, and which single one I should set up first.
Code – It runs without you
Level 3

Code

– It runs without you

The agent stops doing the work and builds the tooling that does it – running on its own, showing you the picture, and reaching your customers.

The shift here is easy to miss, because it sounds like more of the same. Until now the AI has been doing your task. At this level it stops doing the task and builds the thing that does it – and that thing keeps working while you are asleep, on holiday, or busy with something that genuinely needs you.

What it unlocks

The daily and weekly work becomes software you own. Dashboards for the parts of the business you keep asking about, and real experiences for your customers.

Where it runs out

It takes practice – first version to something reliable is real work.

Unattended

The work happens on a schedule, when you are not there. Friday afternoon, seven in the morning, the first of the month.

Visible

A dashboard for the part of the business you keep asking about – leads, traffic, campaigns, operations. Consistent in a way a chat answer never is.

Customer-facing

Real products. Onboarding that runs itself, lead-generation experiences, the small tool your customers use and remember you for.

Only worth running once a task has genuinely worked at Level 2 for a few weeks. The questions it asks first are the ones people skip and then regret.

I've been doing [task] with your help for a few weeks now and it works well enough that I want it to run without me.

Before proposing anything, ask me: how often it should run, what it should do when something looks wrong, who else needs to see the result, and what it must never do on its own.

Then describe the smallest version that could run once, unattended, this week. Not the full system – the first honest version of it.

No, You Don’t
Have to Start Over

Here is the objection almost nobody says out loud, and it stops more people than any technical difficulty ever has: if I go past chat, I have to learn a whole new thing, and I have already put the work into getting good at this one.

You do not switch tools to move up. The same Claude Desktop you chat in at Level 1 gains connectors and skills at Level 2, then opens Claude Code at Level 3. Nothing to start over.

Sit with that, because it inverts how this is usually sold to you. There is no migration. Your habits stay exactly as they are and gain reach. The window you already type into is where all three levels happen.

There is another road, and it would be dishonest not to name it. You can reach Level 3 through a chat builder like Lovable or Base44 instead. It is faster to a first screen and it leaves you somewhere else – a separate tool, holding an app you cannot take further.For a first screen or a weekend prototype that trade is often the right one – I have made it myself and published the whole build log. Just choose it knowing what it costs, rather than because it was the first thing you found.

Not Everyone Needs Level 3.
Everyone Has a Minimum.

Climbing to the top is not the goal. Knowing where your floor is – for the way you actually make money – is. Below it the cost is real but quiet. It shows up as margin, as speed, as the gap between what you can promise and what you can deliver, and never as a single bad day you could point at.

Read the third column as a description of a trade-off, not a warning. Plenty of people sit below their floor on purpose and do fine there.

If you areYour floorWhat staying below it costs
Service business or consultingLevel 2 · Co-WorkYou scale by adding hours. Client tools and automated delivery stay out of reach, so a one-person shop can never look like a team.
Agency or studioLevel 2 · Co-WorkEvery deliverable eats margin while the tooling stays manual. Growth means hiring.
Software product or SaaSLevel 3 · CodeProduction breaks demand your full attention and nothing runs unless you trigger it. You end up competing on manual execution speed.
Content, courses or a communityLevel 2 · Co-WorkContent gets produced by hand at a pace that cannot compound. The distribution machine never gets built.
EmployedLevel 2 · Co-WorkYour team stays dependent on central IT and outside vendors, on their timeline instead of yours.
Still deciding what to buildLevel 2 · Co-WorkDeciding without evidence keeps you circling. Co-Work is how you test three ideas in a week instead of one in a quarter.

One row needs Level 3. If you sell software, nothing runs unless you trigger it, and you end up competing on how fast you personally can execute – a race you eventually lose to whoever built the instrument. Everyone else floors at Level 2, and Level 2 is genuinely reachable in a week.

The Same Hour,
Every Month

Everything above is also a live session called AI Compass. One hour, online, free. The room votes on which task we do, and then I run that task up all three levels on screen with real data – including the parts that go wrong, which are the parts you cannot get from a tutorial.

Here is the bit I would normally be advised not to tell you: it runs every month, and it is the same session every time. Same three levels, same structure, same close. No last chance, no cohort closing, no reason to panic about the date.

What changes is the quality. Every session ends with the room telling me which moment landed and what they still could not see, and the next one is built on that. The free-text box in the walker feeds the same list – whatever you typed up there is a candidate for the task we build live next time.

So the honest pitch: come on 26 Augustif it fits your calendar. If it doesn’t, come to the next one, which will be better.

Join AI Compass on 26 AugustOr just start the 7-day plan

One Task,
One Level, Seven Days

This moves exactly one recurring task from Level 1 to Level 2. Not your workflow. Not your stack. Not three tasks. The reliable way to finish the week with nothing moved is to try to move everything.

It is also deliberately unambitious about quality. Day four asks you to do the task badly, on purpose, because what you are testing is whether the connection works at all – and tuning a prompt on top of a connection that is not there is the most common way this week gets wasted.

That is the whole method. One task, one level, and an honest admission that the first attempt will be poor. Do it once and the second one costs you almost nothing – which is the part nobody mentions, and the reason this compounds instead of joining the list of things you tried.

Cheers,
Ben

Ready to Go Deeper?

Pirate Lab

·Wed, Sep 9 · 18:00 CEST

Free weekly online workshop where we walk through the week's Captain's Insight together. Bring your project, get live feedback.

Vibe Coding Cologne

·Wed, Oct 7 · 18:30 CEST

Monthly in-person meetup in Cologne. Talks from Ben and local founders, drinks, building alongside the community.

Vibe Hackathon Cologne

·Fri, Nov 20 · 14:30 CET

Intensive on-site weekend hackathon – build and ship something real in 48 hours with other founders in Cologne.

Pirate Forge

·In 22 days · Wed, Sep 30

6-week cohort program combining build and grow tracks. Weekly workshops, accountability, the founders you'd want to ship next to.

Questions & Answers

Do I need to learn to code for Level 3?
No, and I'm the evidence – I spent years running businesses while depending on technical co-founders to build what I could already sell. What changed wasn't me becoming technical, it was that the tools moved. At Level 3 you describe what you want and review what comes back, in the same conversational way you already work at Level 1. What you do need is patience, because the first version of anything is rough, and the willingness to click through what it built rather than trusting the summary of what it built. That second one catches more problems than any amount of technical knowledge.
Is this safe with company data? What actually leaves my machine?
Be precise about the question, because "is AI safe" has no useful answer. What matters is: which specific place did you grant access to, read-only or read-write, and what is that provider's retention policy. A connector is scoped – granting calendar access does not grant inbox access. Business and enterprise tiers of the major assistants contractually exclude your data from training; consumer tiers often do not, and that difference is worth ten minutes of reading before you connect anything real. My own rule is boring and has served me well: start with a connector that can only read, and only add write access once you have watched it work for a week.
My company won't let me connect anything. What can I still do?
More than you'd think, and this is a common situation rather than a dead end. Exported files in a folder get you most of Level 2's value – it's clumsier, but the AI is still working from your real information rather than guessing. Start there and you'll have something concrete to show, which is a far better argument with IT than a request based on an article. In my experience the teams that get connectors approved are the ones who arrive with a working example and a specific scope, not the ones who ask for general permission.
How much does this cost per month, realistically?
Level 1 and most of Level 2 sit in the twenty-to-thirty a month range per person for a business tier of a major assistant. Level 3 typically means a coding agent on top, which is where the number moves – budget somewhere in the low hundreds if you're building regularly, less if you're not. The honest framing is that the cost is real but it's almost never the constraint. The constraint is the hour you have to spend on day three connecting the first tool, and no subscription fixes that.
What if my task turns out to be a bad fit – how do I tell early?
Day four tells you. If it could not reach the information it needed, the task depends on something that isn't accessible – a system with no connector, a colleague's head, a paper form – and no amount of prompting will fix that. Either solve the access problem or pick a different task. The other early signal is a task where the hard part is a judgement call only you can make. AI can prepare that decision beautifully and cannot make it, so those tasks get faster but never leave your desk.
Should I use Lovable or Base44 instead?
For a first screen or a weekend prototype, honestly yes – they're faster to something you can look at, and I've built in Base44 myself and published the whole thing. The trade-off is where you end up: a separate tool, holding an app you can't easily take further, with the platform owning the hosting and the auth. That's a fine bet when the thing is small and self-contained. It's a worse bet when the tool is going to live inside your business for years. The reason I recommend staying in one environment is continuity, not capability – both roads reach Level 3.
How long does Level 1 to Level 2 actually take?
A week for one task, which is exactly what the plan above is scoped to. Getting comfortable enough that you reach for Level 2 by default takes more like two months, and it happens task by task rather than as a single moment. The people who stall are almost always the ones who tried to move their whole way of working at once instead of one recurring job.
I'm not a founder, I'm employed. Does this apply?
It applies more directly, if anything. Your floor is Level 2 for the same reason a consultancy's is – below it, your team stays dependent on central IT and outside vendors, on their timeline rather than yours. The person on a team who can connect AI to the actual work becomes the one who unblocks things, and that is visible in a way that generic AI literacy never is. The scope is smaller than a founder's, which usually makes it easier: you have one clearly annoying recurring task and you don't need anyone's permission to describe it.
What if I'm already at Level 2 – do I need Level 3?
Only if you're selling software, where it isn't optional. For everyone else it's a genuine choice, and the test is repetition plus consistency: if you find yourself asking for the same thing weekly and wanting the answer to look the same every time, that's an instrument waiting to be built. If your Level 2 tasks are varied and each one benefits from you steering it, you're getting most of the value already and Level 3 would mostly buy you maintenance work.
When is the answer "don't bother"?
A task you do twice a year is not worth handing over. Neither is one that takes four minutes, however irritating those four minutes are – the setup will cost more than the task ever will. And a task whose difficulty is entirely a judgement call is one AI should prepare and never own. Say no to those quickly. The reason this matters is that people often pick their most annoying task rather than their most repeated one, and annoyance is a poor proxy for frequency.
Does it matter which AI tool I start with?
Much less than the internet suggests. The frontier models are close enough that your choice of task matters more than your choice of assistant. The one thing worth checking before you commit is whether it connects to the tools your work actually lives in, because that's the whole of Level 2 – a model that scores better on benchmarks but can't see your calendar is the wrong tool for this. I use Claude, which is a preference, not a recommendation you need to follow.
What breaks most often at Level 3?
Two things, and neither is the code. The first is that something upstream changes – a document gets renamed, a field gets added, an integration quietly expires – and the tool keeps running while producing nonsense. Which is why anything running unattended needs to tell you when it did nothing, not only when it did something. The second is trusting the report over the reality: it will tell you a feature is finished and working when clicking through it finds a dead end. Click through it yourself. Every time.
Ben Sufiani, The Captain

Ben Sufiani

The Captain

Founder from Cologne with 15 years of startup experience across 9 ventures. After helping thousands master growth marketing, Ben learned vibe coding from scratch and launched CaptAIn within three months. He leads the Vibe Coding Cologne community, blending real founder experience with teaching clarity.