The Spec: how we plan our software projects in the age of AI

AI is fundamentally changing software development right now. We see it as our job to make that change work for you, our clients. So you can benefit from faster results and more transparency.

AI coding tools are no longer a gimmick for us. We use them every day, and that’s changing how we build and think about software.

Just one figure from our day-to-day: roughly 50% of our code is currently AI-generated. The rest? Review, architecture, conception. The work that makes sure the result actually works.

And this is just the beginning. Gartner forecasts that by 2028, 90% of all software developers will be using AI code assistants (compared to 14% at the start of 2024). Microsoft CTO Kevin Scott goes further: he expects 95% of code to be AI-generated within the next five years. Personally, I share that view.

The problem nobody likes to talk about

‘Vibe coding’, the practice of quickly clicking code together with AI support, works surprisingly well. Until it doesn’t.

Sooner or later, every AI-driven project hits its limits. The codebase becomes so complex that the AI gets in its own way. Without clear structure, without documented architecture decisions, without traceable requirements, it produces code that runs in the short term but turns into a maintenance nightmare in the long term.

The question is no longer whether AI writes code. The central questions are: does this code have reliable quality? Will it still be extensible in two years? Does anyone still understand what the system is supposed to do?

What’s shifting fundamentally right now

We’re seeing three developments:

The value of ‘typing code’ is dropping rapidly.

What used to take weeks, an AI agent now does in hours. That’s great, but it also means ‘I can code’ alone is no longer a differentiator.

The value of ‘knowing what to build’ is rising.

Meaningful business cases, clear requirements, thoughtful user flows, precise acceptance criteria: that’s the new bottleneck. Not the typing, but the thinking that comes before it.

The value of architectural knowledge stays high.

Someone needs to know what good software architecture looks like. How to build a system that scales, stays maintainable and can be extended. AI can’t do that (yet).

The conclusion for us: we need a new way to specify projects. One that’s understandable for humans and that AI agents can use as a working basis.

We call it ‘The Spec’.

What is the Spec?

The short version: The Spec is structured project documentation with three properties:

• Single source of truth: everything the system should do is in one place
• Lives in the code repository: not in a wiki, not in JIRA, but right next to the code
• AI-readable: structured so that AI agents can work with it

The Spec contains two things: project-specific requirements (what should this system do?) and our technical best practices and architecture standards (how do we generally build good software?).

‘But isn’t that just like a specification document?’

Fair question. The reason classic requirements and specification documents have fallen out of use: the maintenance burden was enormous. Every change meant hours of documentation work, and in the end, the document was outdated anyway.

With us, it’s different. The Spec is machine-readable, and AI agents can support change updates. Changes are part of the normal flow, not a separate documentation process. That was the actual problem with the old specifications: every change was expensive. With us, changing is cheap, and we stay agile.

The three layers of the Spec

The Spec is organised in layers. The deeper you go, the more technical it gets:

This means: as a project lead or member of management, you don’t need to understand the technical details. You read the first two layers and you know what’s being built. The developers go deeper.

What this means for project management

For project managers, the Spec solves concrete everyday problems:

What this means for you

Efficiency: more output for the same investment.

We get faster. Not because we cut corners, but because routine work is automated. Features that used to take us two weeks, we now do in one. Or we use the time we’ve gained to test the feature with users and iteratively improve it based on user feedback.

Sovereignty: real independence.

The Spec is a piece of risk management. It protects you at the most vulnerable point of every software development project: the handover. No being left in the lurch when teams or suppliers change.

Iteration speed.

The 2-week sprint rhythm fits most projects. But when a shorter time-to-market or an intensive product development cycle is needed, we can pick up the pace.

How humans and AI work together

Our goal is a clear division of labour:

Humans stay in charge. But the routine work (writing standard code, running tests, updating documentation) moves to AI.

Why the Spec is decisive for this: AI agents are only as good as their context. If we tell an AI agent ‘Build me a login’, we get just any login. If we give it a Spec with acceptance criteria, security requirements, UI design and data model, we get the login we need.

The Spec contains our expert knowledge, defines clear boundaries and gives AI agents precise requirements.

Which projects is the Spec right for?

Worthwhile for projects of 6+ months.

For projects that should live longer, that will be evolved, that may at some point be handed over: that’s where the Spec pays off.

Overkill for throwaway MVPs.

If you want to know in two months whether an idea works and the result is going in the bin anyway, you don’t need a Spec.

More care upfront.

We need to know more precisely what we’re building during the conception phase. That requires better workshops, clearer requirements and more thoughtful designs. That effort pays back many times over across the project lifetime.

What’s next?

We continue to engage intensively with these changes, experiment daily, and we’re convinced this is the right path.

Our concrete plan for 2026:

• Q1/Q2: rollout on selected new projects
• Ongoing: iterative improvement based on real-world experience
• In parallel: expanding our AI agent infrastructure

Want to find out whether spec-driven development is right for your project? Get in touch with me for a personal conversation. No obligations, no sales pitch, promised. I’m looking forward to it!