AX Framework Starter Kit
How to start designing AI with intention
Explained Step-by-Step Through a Practical Case Study
AI is moving into products faster than teams can ask the right questions. It’s common to see projects start from the wrong end: exploring models, tuning prompts, building impressive demos… and only later trying to understand how the AI should actually behave, what boundaries it needs, and what impact it will have on real people.
The AX Framework Starter Kit exists to flip this sequence. It’s not the full AX Framework; it’s the entry point. A simple, accessible, and intentional way for teams to begin designing AI systems that are responsible, understandable, and trustworthy. The logic is straightforward: before the technology comes the human; before capability comes intention; before output comes behaviour.
To illustrate how it works in practice, we’ll walk through a concrete, relatable example - HomeMeal, a small feature that demonstrates how improved AI behavior design can transform the entire experience.
The Case: HomeMeal
Imagine HomeMeal, an app that suggests meals based on the ingredients the user already has at home. It’s not a chatbot. It doesn’t “talk”. It’s simply part of a daily routine: open the app, check the inventory, and get quick, realistic, and safe suggestions.
The problem is familiar: overly complex recipes on days when the user is exhausted, suggestions that ignore allergies, combinations that assume ingredients that don’t exist, and explanations that are too vague to be helpful. The AI starts as support… and quickly becomes noise.
The Starter Kit helps turn this around. Let’s go step by step.
1. Start with the Human Context
The AX method begins with a simple truth:
You cannot design intelligent behaviour without understanding human behaviour.
The starting point is the Tri-Canvas, available at:
https://axframework.portoux.com/tri-canvas
A HomeMeal user arrives home tired, with limited time, an imperfect inventory, and dietary considerations. The AI needs to reduce cognitive load, not add to it. It needs to be practical, predictable, and respectful of the user’s real-life constraints.
Tri-Canvas - People
The user is tired, wants something quick, and cannot risk a dietary mistake. The AI must prioritise simplicity, time awareness, and safety.
2. Define the Role and Boundaries of the AI
The second step of the Tri-Canvas clarifies what the AI can do and what it should never do.
Tri-Canvas - Machine
The system recognises expiry dates, preferences, allergies, and available time. It organises this information so the AI can make grounded, realistic suggestions.
Tri-Canvas - Intelligence
The AI cross-references ingredients with validated recipes, prioritises items that are about to expire, adapts complexity to context, explains the reasoning behind each option, and asks for clarification whenever data is unclear. When risk exists, it clearly alerts the user. When unsure, it asks instead of guessing.
HAX patterns (from Christian Kuhn) reinforce this layer, prompting the AI to explain before acting, request clarification when needed, and reveal uncertainty honestly.
3. Design the AI’s Behaviour
Behaviour design happens in the AX Sprint Canvas:
https://axframework.portoux.com/ax-sprint-canvas
Get Marta Fernandez’s stories in your inbox
Join Medium for free to get updates from this writer.
This is the moment when HomeMeal stops being an idea and becomes an intentional behavioural system. Below is the filled example exactly as it would appear inside the AX Sprint Canvas.
AX Sprint Canvas - HomeMeal example
Problem
People reach the end of the day exhausted and overwhelmed by the small but heavy decision of “What should I cook?”.
This leads to repetitive meals, poor choices, or wasted ingredients.
User Scenario
Evening context. Low energy. Limited time. An imperfect inventory. Dietary restrictions and preferences that must be respected. A need for simplicity, clarity, and safety.
AI Role
The AI acts as a decision-support system, not a decision-maker.
Its purpose is to reduce cognitive load, propose realistic meal options, and adapt to the user’s constraints without overriding personal judgment.
Boundaries
The AI must not invent ingredients, ignore allergies, suggest unrealistic or time-consuming recipes, or act with unjustified confidence.
If information is missing or inconsistent, it must be asked for before suggesting.
Required Inputs
Inventory list, expiry dates, dietary restrictions, user preferences, available cooking time, and usual kitchen equipment.
Expected Behaviour
The AI identifies items nearing expiry, cross-checks them with suitable recipes, adapts complexity to the time available, and explains why each suggestion is relevant.
When uncertainty appears, it asks the user to confirm.
When confidence is low, it chooses safer and simpler alternatives
4. Validate, Launch Carefully, and Monitor Continuously
An AI can behave well today and drift tomorrow.
This is why the AX Framework includes a final layer of continuous evaluation.
The AX Evals Kit is available at:
https://axframework.portoux.com/resources
For HomeMeal, only the three mandatory sections apply:
Behaviour Integrity, Trust & Interaction, and Evolution & Drift.
AX Evals Kit - HomeMeal example (mandatory sections only)
Behaviour Integrity
We verify whether the AI behaves exactly as designed.
In HomeMeal, the AI respects boundaries, asks for clarification when inventory data is ambiguous, expresses uncertainty when needed, and explains each choice with enough context. It is consistent: similar situations lead to similar suggestions. It does not invent ingredients or expand into unauthorised behaviours.
Trust & Interaction
We check whether the user understands the suggestion, feels in control, and maintains a healthy level of trust.
In HomeMeal, explanations are simple (“I suggested this because your mushrooms expire today and you have 20 minutes”), and the user can reject suggestions or ask for alternatives without friction. Trust is balanced, neither blind nor skeptical. The AI reduces effort instead of adding to it.
Evolution & Drift
We ensure the AI remains stable over time.
In HomeMeal, the AI doesn’t fall into repetitive loops, doesn’t ignore dietary restrictions, and doesn’t lose diversity in its suggestions. Explanations remain clear and consistent. Trust levels are stable, and the team adjusts behaviour intentionally based on feedback. There are no signs of behavioural drift.
Wrap-up: the Starter Kit is the beginning, not the destination
HomeMeal shows how the AX Framework Starter Kit can transform a simple feature into an intelligent system that acts with intention, explains its reasoning, respects limits, and remains stable over time. The AI stops being a generator of outputs and becomes a reliable partner, because it was designed with attention to what truly matters: the human context.
But this is only the surface.
The Starter Kit gives teams the essentials, the minimum needed to avoid the common mistakes of AI-first work and begin building systems responsibly. Its strength lies in unlocking clarity and focus at a moment when many rush into models and prompts without defining how the AI should actually behave.
The full AX Framework goes much further.
It brings ethical governance, continuous monitoring, behavioural stability, advanced explainability, co-evolution, proportional risk management, and the unifying principles that align entire teams, design, engineering, product, and security, around a shared language for human-centered AI.
The Starter Kit opens the door.
The full AX Framework builds the path.
For teams striving to design AI that deserves to exist, AI that people understand, trust, and choose to use, the AX Framework isn’t just useful. It’s essential.
If you’d like to explore the AX Framework beyond this Starter Kit or understand how to apply it to your own AI features, you’re welcome to reach out. Whether you need deeper clarification, want to test the framework on a real product, or are considering a workshop for your team, feel free to contact me.
Designing AI with intention is a journey, and no one should navigate it alone.
