Generative Storytelling

AI for ADHD

devlin.bentley@gmail.com (Devlin) — Wed, 15 Apr 2026 16:44:39 -0700

Let’s Fix ADHD For Knowledge Workers

There aren’t any good accommodations for people with ADHD. Employees are stuck with egg timers at best,and employers are stuck w/o a way to help.

Let’s fix that.

Disclaimer

ADHD presents differently in different people. What works below won’t help everyone, but holy cow is it better than the current state of the art.

Understanding Motivation

The Fogg Behavior Model | Behavior Design Lab is wonderful for understanding what gets people to do things. Look at this wonderful graph.

For those who haven’t attended one of BJ Fogg’s wonderful training courses (they are legitimately amazing), the idea here is that people who are really motivated can do really hard things, but the less motivation people have the less difficult tasks they can take on. We can use “prompts” to make things happen and build habits, but that is a separate topic.

One symptom of ADHD is that it makes things appear harder to do than they really are. Another symptom is fast fall off of motivation, and easily being discouraged. Basically that green line becomes artificially distorted.

A key takeaway is that ADHD isn’t a motivation problem. People with ADHD have just as much desire to get things done as anyone else. The difference is, if you tell someone with ADHD that washing their car first involves finding a bucket, mop, some dish towels, and soap, they are going to feel overwhelmed and choose to do something else. This is why organizing one’s environment is so important with ADHD. If a person with ADHD wants to regularly clean their car, they are going to keep all of the car cleaning stuff together, even if that means buying extra dish towels just to keep with a dedicated mop and bucket for car washing. Now “get ready to clean the car” becomes two steps - grab the car cleaning kit, add water.

Another example - someone with ADHD who likes to cook may choose to alphabetize their spice rack. This can dramatically reduce the mental effort to cook complicated dishes.

This is the key - minimize unneeded mental “prep work”, streamline, simplify, reduce the number of steps needed to get started on a task.

Let’s AI The Hell Out Of This

Time to throw some local models at this.

First off, the interaction modality is always listening voice. Remove even needing to remember another hot key. Keyword activation is the name of the game.

The assistant does the following:

Sets timers for work/break
Reminds them of upcoming meetings and any prep work that is needed before hand
Maintains a Rolodex of people, names, jobs, etc
Tracks to-dos

A more advanced form:

Goes through email and Slack message, summarizes anything that needs attention
Maintains notes about particular projects or topics

Now, one by one.

Timers

Timers are great. Timers are the #1 reason why people use Alexa. They aren’t generally monetizable in a SaSS sense, despite tons of efforts to try. They need to be easy to set, which is why everyone loves voice timers. They are a must have for any productivity solution. Pomodoro timers work really well for many people and they are easy to add, so add them.

Meeting Prep

People manually take notes in Obsidian. Notion automatically takes notes. Zoom automatically takes notes. None of those notes are used for anything. Before every meeting the AI should gather a list of who is in the meeting, do a recap of previous interactions with those people, and prepare links to any needed documents for review. This functionality should be accessible at any time during the day. “What documents do I need to review for my meetings today?”

Rolodex

After a meeting it should be possible to go “Hey, DesktopAI, note down that Joel is the expert on our security review process.”

2 months later - “Hey DesktopAI, who was the person I spoke to awhile back about our security reviews?”

No more going back through the calendar determining who was in a meeting 2 months ago. Ideally this would be tracked automatically based on meeting transcripts, but baby steps for now.

To-dos

“Hey DesktopAI, add schedule a security review with Joel to my to-dos for tomorrow”.

“Hey DesktopAI, summarize my to-dos for this morning by priority.”

“Hey DesktopAI, go through my Slack channels and find anything needing my immediate attention within the next 2 hours, summarize, and add to my to-dos.”

“Hey DesktopAI, I scheduled that review with Joel, mark it as done.”

Integrate with Linear/Jira as desired.

Slack/Email integrations

Remember when Gmail used to do this, like a decade ago? Wow it was nice.

Again, simple enough. There is so much garbage in my inboxes, let me know what actually needs my attention.

Topic Notes

In the post Useful Corporate Agents I discuss how this should work in depth. The short of it is, there should be a way to have notes on a per topic basis that can be referenced. This keeps life organized in a meaningful way.

“Hey DesktopAI, add a note to the front-end rework project that we need to decide on a CSS framework.”

“Hey DesktopAI, for the vendor contracting process, add a note that Xyzg systems is in the running now.”

It should be trivial to pull up all notes about a project. Bonus points if transcriptions from meetings about a project are all automatically summarized and pulled in as well.

A Recap Of What We’ve Fixed

Getting ready for a meeting -

Before: Open up the invite, look at who is in it, go through description for any needed docs. Go through previous meeting notes to see what was discussed last time. Check manual rolodex to remember who everyone is and what their specialty is.

After: You are presented with a list of docs to review, a summary of all previous related meetings, and a brief description of everyone who will be attending. Steps reduced down to just reviewing linked docs.

Staying Focused At Work -

Before - Manually start a pomodoro timer up on the computer. First install the app, then remember what it is called. After rebooting, find it again.

After - No need to remember an app name, just say start a timer. Again, this sounds simple, but the absurd usage #s behind Alexa timers tells a different story.

Before - Manually write everything down on a to-do list and keep portions of it in sync with your workplace’s Linear/Jira board (as needed), and other portions to yourself (emails to write, meetings to schedule, etc)

After - A list is maintained for you, always accessible, always available, and items can be added w/o interrupting flow.

Finding a Subject Matter Expert

Before - Go through meeting notes manually trying to remember who it was you chatted to a month or two ago

After - Just ask your AI.

How To Build It

SaSS or non-SaSS? Locally hosted or not? Macs with 24GB+ of RAM can do all of this with the latest Gemma 4 models running locally. TTS models are now under 1GB of VRAM, pick your favorite STT (ASR) model.

Cloud based, hook this all up to your favorite provider. I’d still do local TTS to save yourself a ton of money (the CPU only models run just fine now days), but feel free to use whatever -nano variant you want of a cloud based LLM. None of the above requires anything remotely resembling a frontier model.

Build It, Then What?

Then you get it certified as an accommodation for people with ADHD and you sell it to employers, preferably through insurance companies. Since you’ll have the entire market to yourself, enjoy massive profits.

A Vision For Useful Company Wide Agents

devlin.bentley@gmail.com (Devlin) — Wed, 15 Apr 2026 15:30:00 -0700

A Vision For Useful Company Wide Agents

Current business focused agents suck. Some are more useful than others (Notion), some try to be useful (Claude’s business offering) but none of them is holistic.

Current agents can search across multiple sources of data. A few can take some actions, moving linear tickets around, even writing code and committing code.

Your company probably has a transcription agent listening to every meeting, shoving a transcript into a folder somewhere for it to never be looked at again.

You likely have multiple agents running through Slack fetching information on demand, or handling some repetitive tasks.

Your agents should be doing a lot more.

Your Company’s Knowledge Graph

Imagine this world - every meeting transcript is gone through and what people are working on gets tagged, even better what people know about gets tagged. What people are good at is tracked. When someone needs to know who is the subject matter expert on a topic, your company’s AI instantly knows.

Duplicate work or common complaints are automatically tagged and identified. A problem that is happening across multiple teams is brought to attention right away.

From an individual’s perspective, all Slack conversations that are relevant to me are highlighted each morning when I sit down at my desk. I can ask my computer what I need to get started on for the day and I’m given a list of what needs my attention.

Work streams across Slack, email, Linear, Github issues, PR comments, Notion threads, asks for feedback in Figma, are all brought together into one holistic view of what needs to get done.

As we progress into an agentic world, what is most important for people is knowing what needs their attention. Where do I need to focus to oversee what is going on right now?

Pulling This Off

The technology is actually there. Heck this can be built in one of two ways - a centralized solution works, but local models are powerful enough now that Gemma 4 or Qwen 3.5 can do it all locally. Setup MCP servers to Slack and Notion and you are good to go. Just use your favorite meeting recording tool that does diarization.

Local first preserves privacy and is a lot less creepy (an important point to consider for AI products right now). Run a local graph DB, every coworker has an entry, links to all meetings you’ve had with them, information on what they are working on, what topics they are most knowledgeable about. The next type of nodes in this DB are topic nodes, areas of work within the company. Categories are automatically created, regular cleanup tasks consolidate tags and keep the DB clean. Periodically an agent goes through Slack channels, email, Linear, and so on, pulling new content and prioritizing what needs attention.

Local lets this scale well. You really do want each person to have their own view of the world, and this is one of those problems that scales O(n²) with the number of users. Doing this work on the edge saves a lot of resources, but it limits your target audience. An initial product launch would probably be Mac only.

Of course this can be centralized, and that makes it more palatable to investors. I’ve previously written about how a focus on SaSS has introduced myopia across our industry, and this may be another example, but one does what one must to build a viable business.

A Day In The Life

I sit down at my computer at login. CompanyAI lets me know there are 2 new surveys I need to fill out for my team in a manager’s Slack channel. I am reminded of 2 new code reviews seeking my approval, and one of my employee’s has a code review that needs approval from a third party team and he’s been waiting 2 days so I may want to intervene. In a meeting yesterday another team discussed new work items that will impact us, and I’m given the name of the dev leading the effort so I can reach out and discuss the work she’s doing.

Mid day, those items completed, I ask CompanyAI if there is anything else I should be focusing on. I am told the Figmas for our team just got updated and I may want to take a look. Using STT (ASR) I am able to easily add items to my own to do list. Optionally I can create Linear tickets for those to do items that need to be tracked. I’ve been thinking about some of our data normalization needs lately and I ask my Company AI to create me a new spike in Linear, assign it to me, create a new page in my team’s Notion, and link the Linear ticket to the Notion page, and add the spike to my to do list for tomorrow.

After each meeting I attend through the day I am able to quickly tell my CompanyAI what takeaways are most important to me, and also add any to do items to my personal list using just my voice. I can also specifically add new tags or categories that may have come up during the meeting, to help the categorization system with its job.

Throughout the day, my attention is focused and purposeful. When distractions do come up, I can quickly prioritize them for later if need be, and know they won’t be lost. When I am done with a distraction I can just ask my computer what I was doing, and then jump right back into it.

Focus and attention are the most valuable resources we have, let’s use AI to amplify them.

Testing webhooks

devlin.bentley@gmail.com (Devlin) — Mon, 13 Jan 2025 23:37:42 -0800

Testing if web hooks work.

My First Post

devlin.bentley@gmail.com (Devlin) — Sat, 11 Jan 2025 23:37:42 -0800

My first post