The house rules
Architecture, standards, security and regulation, written so the AI applies them every time.
AI-Accelerated Development
We build two things with your team that stay with you: the context the AI works in, and the process you work with, from idea to up-to-date documentation.
Your team defines the feature and adjusts it; the context and the automated checks carry the rest.
We help your development team work with AI on your code. It is for companies with 5 to 1,000 developers, and we start with one team. Quick-start: 4–8 weeks plus 3 months of support. The method and the AI's context stay with your team, and we measure delivery and quality before and after. Start with a free diagnostic.
The method
The ten-step flow in which a development team works with AI without losing control: the context and the automated checks carry the rest.
With AI, writing code is now the cheap part; the hard part is controlling what gets built.
The method is based on Spec-Driven Development (SDD): every change starts from a specification and leaves a trail.
The flow goes from the idea to the feature, the user story, the PRD, the plan, the tasks and the implementation. Agents review and a person approves the change. When the sprint closes, specialised agents audit what was delivered against what was specified, and the documentation is updated with every story.
A person decides where the risk is: the feature, the PRD, the plan and the review. Approving everything is not control, it is a jam.
It starts with a free diagnostic, then a Quick-start with one team (4–8 weeks and 3 months of support) or the full Programme (12–16 weeks and 6 months of support).
Spanish subtitles (the on-screen text of the video is in Spanish).
The context
Coding agents are only as good as what they know about your company. That is why we set up the team's harness with them, the context the AI works in.
Architecture, standards, security and regulation, written so the AI applies them every time.
Modules, decisions and documentation, so the AI starts from what the team already knows.
One for each phase: specify, plan, implement, review and audit.
They stop what does not comply before it reaches production.
We start from onext's own accelerators, configured on your technologies.
That harness stays with the team: it is theirs, and they grow it with every project.
The process in ten steps
With AI, writing code is the easy part; the hard part is controlling what gets built. The team works with Spec-Driven Development (SDD): the specification, not the generated code, is the source of truth.
Idea
The need or opportunity is captured, with no solution yet.
Feature
We define what will be built and why. A person approves it.
A person decides: Approves.
User story
It is written from the user's point of view, with acceptance criteria that can be checked.
PRD (product requirements document)
The requirements document sets the scope, the edge cases and what is out. A person approves it.
A person decides: Approves.
Plan
We decide how it will be built: design, dependencies and risks. A person approves it before any code.
A person decides: Approves.
Tasks
The plan is broken down into small, checkable tasks.
Implementation
Agents build each task under automated checks.
Review
Specialised agents review each story, and a person approves the change before it is merged.
A person decides: Approves the change.
Sprint audit
At the end of the sprint, specialised agents audit what was delivered against what was specified.
Updated documentation
Each module's documentation is updated with every story (a person approves the change), and the sprint closes with its archive and release notes.
Your team defines the feature and adjusts it; the context and the automated checks carry the rest.
How we build them
You roll it out with a pilot team, without stopping deliveries. After a free diagnostic, we set up with the team the context (harness) the AI works in. The pilot then applies the ten-step flow to real work, in parallel with its sprint. Three months of support follow.
Phase 1
A conversation about how your team works with AI today and what we will measure. We set the baseline for delivery and quality: lead time to production, failures after deploying, rework and time in review. If we see no difference, there is no Quick-start.
Phase 2
We set up the team's harness with them: rules, product knowledge, agents and checks.
Phase 3
One team works the ten-step flow on real work it had already planned, in parallel with its sprint.
Phase 4
We support the team for three months. We follow the baseline measures and compare them at the end.
Quick-start · 4–8 weeks and 3 months of support · fixed price per team
When we finish
This is what we build with your team:
The harness: rules, product knowledge, agents and checks, written and maintained by the team.
The ten steps, who approves what, and how documentation is updated with every story.
What we build stays with your team.
With your team
We work with your team, on your code and on real work it had already planned. We do not stop deliveries: a pilot team works alongside its own sprint.
We start with one team and your codebase, not a sample project. The context and the process are built with your architecture, your rules and your backlog.
In the free diagnostic we look at how the team works with AI today: people, processes, tools and indicators. We leave knowing where speed or control is being lost and what the Quick-start would change.
Claude Code, Cursor or Copilot: the method works with any of them. We configure the agents and their rules with the team, and the team learns to use them working on its own work.
Every user story is written with acceptance criteria that can be checked, and automated checks stop whatever does not comply before production.
A person approves the change before it is integrated, without reviewing everything at the same depth: approving everything is not control, it is a bottleneck.
We set the delivery and quality baseline in the diagnostic and compare it at the end of the three months of support, with the team.
What we build, the context the AI works in and the ten-step process, stays with you. It is yours and you grow it with every project. Our goal is that you stop needing us.
Results
Results of the method in product teams; the ×8 comes from a B2B SaaS case.
×8
Functionality delivered per sprint
0
Lost sprints while rolling out the method
−50%
Time-to-production
Every team starts from a different place: in the diagnostic we set the delivery and quality baseline and compare it at the end. We don't promise a number before measuring.
For consultancies and software factories
It is for companies with 5 to 1,000 developers, consultancies and software factories included. We start with one team or one specific service line.
The context the AI works in and the ten-step process are yours: you grow them with every project. Our goal is for you to stop needing us.
The method works with Claude Code, Cursor or Copilot. In each project we choose the right technology and model for your case, and you can change them.
In the free diagnostic we set the delivery and quality baseline with the team and compare it at the end of the three months of support.
We work natively with Claude as our reference model, for technical reasons and not commercial lock-in, and our Quick-start is built on Claude Code. The method works just as well with Cursor or Copilot and we do not tie you to one vendor: what we build with your team, the context and the process, is yours.
Reference
We are helping a public sector development team work with specification-guided AI.
Frequently asked questions
onext, an AI boutique based in Barcelona, supports development teams at companies with 5 to 1,000 developers through AI-Accelerated Development. A team works with us for a few weeks on its own code: we look at how you develop, set up the context the AI works in and the working process, and measure delivery and quality. It starts with a free diagnostic and the method stays with your team.
A course teaches the tool; we work on the team's cycle. With your team we write the rules and context the AI applies on every task and the process from idea to documentation, and we test it on real work. It works with Copilot, Cursor or Claude Code and stays when you change tools. If you only need to learn one tool, a course may be enough.
The Quick-start takes 4 to 8 weeks, plus 3 months of support. The Full Program takes 12 to 16 weeks, plus 6 months of support. You do not need to stop deliveries: we work with a pilot team, on work already planned and alongside its sprint. The free diagnostic beforehand lets us pin down the timeline for your team.
In the diagnostic we set the delivery and quality baseline with the team: time to production, failures after deploying, rework and review time. We track it through the three months of support and compare it at the end.
It depends on the code and on the safety net it already has, and we look at it in the free diagnostic, with the team. The context we set up captures your system's rules, architecture and decisions so the AI starts from what the team already knows. If we see no difference in the diagnostic, there is no Quick-start.
Yes. Our Quick-start is built on Claude Code, and the method works just as well with Cursor or Copilot, without tying you to one vendor.
Two things that stay with you: the context the AI works in (house rules, product knowledge, specialised agents and automated checks) and the ten-step process with four decision points, from idea to up-to-date documentation. They are yours and the team grows them. Our goal is that you stop needing us.
The Quick-start has a fixed price per team, with the scope defined up front: if anything falls outside scope, we discuss it before doing it. We settle it with the team after the free diagnostic. For the Full Program, we also discuss it in the diagnostic.
We measure it from the diagnostic and compare at the end, with the team. We measure delivery and quality, not tool usage. We do not promise a number before measuring.
They are two levels of the same service. The Quick-start (4–8 weeks and 3 months of support) builds the context and the process with a pilot team. The Full Program (12–16 weeks and 6 months of support) redesigns the software development lifecycle. It goes end to end, including the organizational side.
It is development guided by specifications. Before any code, the team writes down what will be built, with criteria that can be checked. Without that, everyone asks the AI in their own way and the result is hard to check. It is the core of the ten-step process.
You already have the tool. With your team we write what surrounds it: the house rules the AI applies every time, the ten-step process and the automated checks. A person decides at four of the ten steps: approving everything is not control, it is a bottleneck. We look at it in the free diagnostic: if the diagnostic shows no difference, there is no Quick-start.
No. SDD does not mean specifying the whole product before starting: each feature starts from its own specification, with acceptance criteria that can be checked, and that specification is approved before the code and lives in the repository next to it. The pilot team works in parallel with its sprint, on work that was already planned, without stopping deliveries.
Yes. It is for companies with 5 to 1,000 developers, consultancies and software factories included. We start with one team or one specific service line. What we build, the context the AI works in and the ten-step process, stays with your team: it is yours and you grow it with every project. The method works with Claude Code, Cursor or Copilot and we don't tie you to a single vendor. We settle the details in the free diagnostic.
The framework behind AI-Accelerated Development, published in the open. Start with the first one if you want to understand why the problem is organizational before it is technical.
A 30-minute conversation about your team.
Request the free diagnosticAnswer 16 questions about how your team works with AI today. It is free.
Start the self-assessmentContact
We look at how your team works with AI today and which context and process would make sense to build first.