Accessible AI Interaction Design

For Audio Developer Conference 2025, I received an accessibility sponsorship. As such I have been thinking about how mental and emotional accessibility is a topic that sometimes gets overlooked.

I really love working with LLMs, but they can have predatory patterns and be very attention-sucking and "yes-sayers."
This can be especially challenging for people with ADHD who might have difficulties in such situations. The constant stream of responses, the lack of pauses, the tendency to agree with everything-it's not great for maintaining focus or critical thinking.

I had previously tried to build a rules-based system where I created 'rules' or 'agent instructions' automatically injected by the llms/tools creator into context window. Funnily enough, they also get distracted and forget to follow or include those rules. (Or maybe they contradict with what they are optimized for). As such, the LLMs are non-deterministic-you can't rely on them to consistently apply instructions, even when you put them there.

Since my general idea was to have it evidence-based, I set out on a journey to design a system which is more deterministic. What I've come to is introducing a layer which the user has control over, which injects into the prompt in a reliable manner. This is pretty simple code, which chooses from one of the rules and injects them into the prompt.

System

The idea is to group the rules so that users can configure which types of rules they like-and even better-to build a system which supports the user on that journey. As such, I've devised a system where it picks a rule at random intervals and injects it into the prompt.

Users can rate how they like certain interactions and their efficacy (though self-rating is obviously flawed, this is what we have for now). The system will then run some simple analysis and let the user know which is recommended and make some proposals for the configuration.

Some examples of rules would be:

• "Pause 30 seconds before you continue"
• "That sounds like a bad idea!" (at random, to support critical thinking)

What's Inside

The repository contains two main components:

Research Modules (17 modules)

I've compiled comprehensive documentation from meta-analyses and reviews for adult ADHD:

• Cognitive Behavioral Fundamentals (modules 01-03): Planning, problem-solving, and psychoeducation
• Emotion Regulation (modules 04-06): DBT skills, mindfulness, and cognitive training
• Workplace & Coaching (modules 07-09): Professional support and neurostimulation
• Extended Interventions (modules 10a-10g): Third-wave CBT, neurofeedback, exercise, sleep, nutrition, digital tools, and HR checklists

Each module includes step-by-step technique protocols, effect sizes, mechanisms explaining why techniques work, traceable sources with page numbers, and practical implementation guidance.

Rule Cards (66 cards)

Structured .mdc files (markdown with yaml headers), designed to be injected into LLM system prompts or context windows. Each card contains metadata (sources, effect ratings), triggers for when to apply the technique, protocols with step-by-step instructions, constraints and safety rules, and metrics for tracking effectiveness.

The cards follow a standardized format (SPEC_RuleCard_v1.mdc) that enables programmatic integration, trigger detection, and outcome logging.

The Vision

The end goal is to enable people to use LLMs in a more self-determined manner and have a counterweight to the attention-sucking, token-lottery which AI companies are inherently incentivized to build. (I recommend looking up Shoshana Zuboff)

I'm really looking forward to continuing this project. I think this is on a natural axis with LLMs general development - the chat-based interfaces work somewhat well, but the next step (next to voice control) will be custom interfaces, and hopefully ones we have agency over! We can already see this in code editors, which let users choose from A/B/C options at some points. For now, OpenAI and Cursor have locked down their core UIs, not allowing extensibility, so we have to do it ourselves.

But I think there's something important here: making LLMs genuinely helpful for people with ADHD, rather than just another platform optimized for engagement at the cost of attention and well-being. Let us be the thought leaders.