đŸ˜¶â€đŸŒ«ïž

you're currently on the personal website of Claudia Richoux

Home

How I Used AI to Outsource My Executive Function

Written by Claudia Richoux

Tags: ai, productivity, adhd, claude

Building a personal operating system with Claude Code when your brain fights you

The Problem: Context Scatter as a Founder

I have ADHD. I run a company. I also have a body that needs PT exercises, meal planning, and regular gym sessions. Every morning I’d wake up to:

  • Linear inbox with issues
  • Slack with @mentions and DMs
  • GitHub with PRs needing review
  • Email (personal AND work)
  • TickTick with half-finished todo items
  • A nutrition plan I’d inevitably forget
  • A workout routine I’d modify based on which body part hurt that day

The cognitive load of just figuring out what to do was exhausting before I’d done anything. Context switching between 8+ apps, each with their own notification system, each requiring different mental models. For someone with executive function challenges, this is death by a thousand cuts.

The Insight: What If AI Could Be My Prefrontal Cortex?

I started using Claude Code for work stuff - code reviews, debugging, writing features. But then I noticed something: Claude was really good at structured decision-making. The exact thing my brain sucks at.

What if I could program Claude to:

  1. Aggregate all my contexts into one place
  2. Enforce good habits through methodology (GTD)
  3. Automate the boring stuff entirely
  4. Guide me through decisions I’d otherwise avoid

The System I Built

/morning - The Daily Aggregator

Every morning, I type /morning and Claude:

  1. Opens my Linear, Graphite, and Slack dashboards
  2. Asks me to paste whatever I see (Claude can read screenshots!)
  3. Fetches via API:
    • GitHub PRs needing my review
    • My PRs that are stuck/approved/need fixes
    • Linear issues assigned to me
    • Email counts from both accounts
    • TickTick tasks due today/overdue
  4. Presents a unified dashboard
  5. Walks me through building today’s plan interactively
## Morning Dashboard

### PRs Needing Your Review (3)
- supplyco#1739 - "Add RLS policies" by @vlad
- supplyco#1725 - "Fix auth flow" by @jannik (CHANGES_REQUESTED by you)

### Your PRs - Status Check
- #1698 "Metrics dashboard" - APPROVED ← merge this?
- #1702 "API refactor" - waiting on reviewers

### Linear Issues (5)
- [SUP-1542] Blocked: dependency on Jannik's PR
- [SUP-1589] Due today: write migration

### Email: Personal 12, Work 8
### TickTick: 4 due today, 2 overdue

Then it asks: “What should go on today’s plan?” and creates TickTick tasks with proper due dates.

The key insight: I don’t have to context-switch. I don’t have to remember which app has what. I just answer questions.

The Hybrid Approach: Dashboards That Won’t Automate

Here’s the thing: not everything has a nice API. Linear has an MCP, but it’s clunky for inbox-style browsing. Slack’s API doesn’t capture the gestalt of “what’s blowing up right now.” Graphite (our PR tool) has no API at all. LinkedIn DMs? Forget it.

So I built a hybrid system. Claude runs:

open "https://linear.app/mycompany/inbox"
open "https://app.graphite.dev/#needs-your-review"
open "https://app.slack.com/client"

Three browser tabs pop up. Claude then asks me to paste what I see - and crucially, Claude can read screenshots. So I:

  1. Cmd+Shift+4 to screenshot the Linear inbox
  2. Paste it into the terminal
  3. Claude parses it: “I see 5 issues assigned to you, 2 are marked urgent
”

Or I just Cmd+A, Cmd+C the text and paste it. Either works.

I’ve also gotten automated scrolling screenshots working with Apple’s screencapture and osascript - it can capture an entire scrollable page. I occasionally use these for long Slack threads or Linear backlogs. Next step is programming those into the workflows so Claude can just grab them automatically.

The key is: my focus immediately returns to the terminal. The browser tabs are open, but I’m not there. I grabbed what I needed and came back. Claude becomes the home base that I always return to after quick excursions into distracting apps.

This matters because my failure mode isn’t “I don’t check Slack.” It’s “I check Slack and then it’s 2 hours later.” The open → grab → return flow keeps the terminal as my anchor.

Why I Could Never Do GTD Before (And Why Claude Changed That)

I’ve read Getting Things Done three times. I’ve tried OmniFocus, Things, Todoist, Notion, paper systems. I understand the methodology intellectually. Inbox zero, weekly reviews, next actions, contexts - I get it.

I just couldn’t do it.

The problem with GTD for ADHD brains is that it requires sustained executive function at exactly the moment you have none. You’re supposed to:

  1. Look at your inbox
  2. Pick ONE item
  3. Ask “is it actionable?”
  4. If yes, ask “what’s the next action?”
  5. Route it appropriately
  6. Repeat

Simple, right? Except my brain does this:

  1. Look at inbox
  2. See 47 items
  3. Feel overwhelmed
  4. Notice an interesting link
  5. Open 12 browser tabs
  6. Forget I was processing inbox
  7. Add “process inbox” to inbox
  8. Close laptop

The methodology isn’t the problem. The execution is the problem.

AskUserQuestion as ADHD Medication

Here’s what /gtd-process actually looks like:

Claude: Processing inbox item 1 of 23:
"interesting article about distributed systems"

Is this actionable?
○ Yes - there's something I need to DO
○ No - it's reference/inspiration
○ Trash - I don't need this

Me: [clicks "No"]

Claude: Where should this go?
○ Someday/Maybe - might read later
○ Reference - save with a tag
○ Trash - actually I don't need it

Me: [clicks "Trash"]

Claude: ✓ Deleted.

Processing item 2 of 23:
"Call dentist"
...

Notice what’s happening:

  • One item at a time. I literally cannot see item 3 until I’ve decided on item 2.
  • Forced binary choices. Not “what do you want to do?” but “pick A, B, or C.”
  • Claude proposes, I dispose. The cognitive load of generating options is removed.
  • Progress is visible. “Item 7 of 23” creates momentum.
  • No escape. I can’t open a new tab because I’m in the terminal and Claude is waiting.

This is the external structure my brain lacks internally.

The “Brainrotted Zoomer” Interface

I call it the “brainrotted zoomer” interface because it’s designed for someone with the attention span of a TikTok scroll:

  • Micro-decisions only. Never “plan your week.” Always “is THIS one thing actionable, yes or no?”
  • Immediate feedback. Task moved, checkmark shown, next item appears.
  • No dead ends. If I say “unsure,” Claude asks a clarifying question. I’m never stuck staring at something.
  • Suggested actions. Instead of “what’s the next action?” Claude says:
What's the next physical action? I suggest:
○ "Call Dr. Smith to schedule appointment"
○ "Text Sarah about Saturday plans"
○ "Research best options online"
○ Other - I'll type it

I can just tap an option. The verb is already there. The specificity is already there. I’m not generating, I’m selecting.

Why This Finally Works

Traditional productivity systems assume you can:

  • Hold context in working memory
  • Generate options from scratch
  • Maintain focus through a list
  • Self-direct without external structure

ADHD means I can’t do any of that reliably. But I can:

  • Answer a direct question
  • Pick from multiple choice
  • Follow along when someone else leads
  • Stay engaged when there’s immediate feedback

Claude becomes the external executive function. It holds the context (“you’re on item 7 of 23, you’ve trashed 3, moved 2 to someday”). It generates the options (“here are three possible next actions”). It maintains focus (“okay, now here’s item 8”). It provides structure (“is this actionable, yes or no?”).

I’m not doing GTD. Claude is doing GTD to me. And that’s why it finally works.

/inbox - Email Triage Without the Dread

Same principle applied to email:

Claude: Email 3 of 18:
From: eric@supplier.com
Subject: Re: MacBook order status
Received: 2 days ago
Preview: "Hey, following up on the laptop order..."

What would you like to do?
○ Tell me more - show full email
○ Archive - I've handled this
○ Create todo - need to respond/act
○ Draft reply - help me write back
○ Skip - deal with later

If I pick “Tell me more,” Claude fetches the full email, summarizes it, and asks again. If I pick “Create todo,” it asks for a due date and creates a TickTick task like “Reply to Eric about MacBook order” with proper GTD formatting.

The inbox goes from “264 unread anxiety pile” to “answer 18 questions and you’re done.”

The Recursive Loop: Everything Flows Back

Here’s where it gets elegant: the system feeds itself.

When I process email with /inbox and choose “Create todo,” that task goes into TickTick. Later, when I run /gtd-process, that same task comes back through Claude for processing. If it’s vague (“deal with MacBook situation”), Claude helps me sharpen it (“Email Eric confirming delivery date”).

If it needs research, I run /research-task on it - which produces more tasks that flow back into TickTick. If those tasks produce emails, those emails get processed in /inbox.

Email inbox → /inbox → TickTick task
                            ↓
                      /gtd-process → sharpen task
                            ↓
                      /research-task → research + new tasks
                            ↓
                      TickTick → more tasks for /gtd-process
                            ↓
                      Action → sends email → response arrives
                            ↓
                      Email inbox → /inbox → ...

Everything flows through Claude. Claude is the processing layer between all my inboxes. The recursive nature means nothing gets lost - it just keeps cycling through until it’s either done or explicitly deferred to Someday/Maybe.

/research-task - When You Don’t Know What To Do

Some tasks sit in my Next Actions list because I don’t actually know how to do them. “Save money on Cytopoint” - my dog needs this expensive medication. I know I should research alternatives but
 that’s vague and overwhelming.

So I run /research-task "Save money on Cytopoint" and Claude launches into research mode:

Claude: I'll help you research cheaper Cytopoint options. First, some questions:

What's your current situation?
○ Getting it from the vet directly
○ Using Chewy/online pharmacy
○ Not sure of current price
○ Other

Me: [clicks "Getting it from vet directly"]

Claude: What's your dog's weight/dosage?

Me: 40lb, monthly injection

Claude: Got it. Let me research options...
[web searches, fetches pharmacy pages, reads rebate programs]

Ten minutes later, Claude produces a memo:

## Research: Cheaper Cytopoint Options

**Current cost:** ~$80-150/month at vet

**Findings:**

1. **Costco Pharmacy** - $45-65 for same injection
   - Has pet Rx program, no membership needed for pharmacy
   - Delivers to home
   - Need vet to send prescription

2. **Zoetis Rewards** - Manufacturer rebate program
   - $30 back after 2nd injection
   - $50 back after 4th injection
   - Stackable with pharmacy savings

3. **Librela** - Newer alternative medication
   - Similar mechanism, may work for some dogs
   - Slightly different side effect profile
   - Worth discussing with vet

**Recommendation:** Switch to Costco pharmacy + enroll in Zoetis Rewards. Ask vet about Librela as backup option.

Then Claude asks what to do with this:

How would you like to proceed?
○ Create next-action tasks (I'll schedule the calls/emails)
○ Make original task actionable (update with specific next step)
○ Just the memo - I'll handle manually

I pick “Create next-action tasks” and Claude creates:

  • “Call Costco pharmacy about pet Rx delivery” - due tomorrow
  • “Email vet requesting Cytopoint prescription to Costco” - due tomorrow
  • “Sign up for Zoetis Rewards rebate program” - due today (quick)
  • “Ask vet about Librela at next appointment” - due next month

The vague “save money” task becomes four concrete actions with verbs and dates. The research is done. The decisions are made. I just execute.

/hevy - AI-Built PT Routines

This is where it gets weird and cool.

I had chronic pain issues. I’d done PT but couldn’t keep track of which exercises to do. So I started having conversations with Claude about my symptoms:

“I have forward head posture and my right hip is always tight. Also my feet collapse inward when I squat.”

Claude would research the biomechanics, suggest exercises, and explain the mechanism. Then I’d say “add that to my workout routine” and Claude would call the Hevy API to update my gym app directly.

My current routines have exercises like:

  • Chin tucks and wall slides (forward head posture)
  • Face pulls with notes: “Pull to face, elbows HIGH. Squeeze rear delts.”
  • Short foot with notes: “Shorten foot WITHOUT curling toes. Pull ball toward heel, create dome. Toes LONG/relaxed.”
  • Dead bugs with notes: “RIB CONTROL: Exhale fully, ribs drop to floor. Low back FLAT.”
  • Tibialis raises with notes: “Back to wall, feet 12in out. Lift toes toward shins. Pause 1s at top, 3s down.”

The form cues are embedded in my workout app. When I’m at the gym, I just open Hevy, and the routine Claude and I debugged together is right there with instructions. No memory required. No “what was that exercise my PT showed me?”

I ran diagnostic conversations like:

  • “My right shoulder clicks when I raise my arm overhead” → added specific rotator cuff work
  • “I can’t get into a deep squat without my heels rising” → added ankle mobility (knee-to-wall) and calf stretches
  • “My lower back hurts after sitting” → added hip flexor stretches, glute activation, core anti-extension work

Each conversation produced updates to my actual workout routine. The gym became a place to execute, not to think.

AWS Lambda + SSM: “Claude 1Password”

My YNAB (budgeting app) was always messy. Amazon transactions would say “AMAZON.COM” with no detail about what I bought. I’d have to cross-reference email receipts manually to remember if that $47.23 was cat food or impulse purchases.

So I built a Lambda that runs every hour:

  1. Fetches Amazon/Apple receipt emails from Gmail
  2. Parses item names and prices from the HTML
  3. Matches to YNAB transactions by amount and date (±5 days tolerance)
  4. Uses Claude Haiku via AWS Bedrock to shorten long product names (40+ chars → 2-4 words)
  5. Updates the YNAB memo field automatically

Now my transactions auto-label themselves:

Amazon: Cat Food 15lb ($24.99), Protein Bars 12ct ($18.99), +2 more
Apple: Claude Pro Subscription ($20.00)

All credentials live in AWS SSM Parameter Store. I call it “Claude 1Password” - Claude can fetch any API key it needs with aws ssm get-parameter, and I never have to paste secrets into chat or worry about them leaking.

The Terraform for this is like 100 lines. EventBridge triggers hourly, Lambda runs, my budget stays accurate without me thinking about it.

Grocery Lists + Nutrition Autopilot

I used to meal plan by staring at an empty document, getting overwhelmed, and ordering DoorDash.

Now:

  1. Tell Claude my macro targets and dietary restrictions
  2. Claude suggests a week of meals
  3. I pick the ones that sound good
  4. Claude generates a grocery list organized by store section
  5. I buy roughly the same stuff every week with minor variations

The decision fatigue is gone. I’m not standing in the grocery store wondering what to buy. I have a list. The list produces meals. The meals hit my macros.

Nutrition is on autopilot. I follow the system.

The Expanding Frontier

The system keeps growing because it’s self-reinforcing:

Learning makeup: “What order do I apply these products? What’s wrong with my technique based on this photo?” Claude can look at selfies and give specific feedback.

Fashion: “I’m going to [X event], I own [Y items], what works together?” Outfit planning without the paralysis.

Company planning: Strategic roadmaps, investor update drafts, hiring rubrics, competitive analysis. Work stuff, but guided instead of blank-page.

The Snowball Effect

Here’s what nobody tells you about integrating AI into your life: it compounds.

The more context Claude has about:

  • My work projects (Linear, GitHub, codebase)
  • My habits (TickTick patterns, what I defer vs complete)
  • My body (PT history, what’s worked)
  • My preferences (past decisions, communication style)

The better its suggestions become. Claude starts to know that I always defer dentist appointments, so it asks “is this actually happening or should we Someday/Maybe it?” Claude knows I respond well to morning workouts, so it doesn’t suggest evening gym sessions.

It’s like having a chief of staff who’s read every email, attended every meeting, and remembers everything. Each new integration makes the whole system smarter.

What “Autopilot” Actually Means

My daily overhead is now:

  1. Morning: Run /morning, answer questions for 5-10 minutes, get a plan
  2. Gym: Open Hevy, follow the routine Claude and I designed. Every exercise has form cues.
  3. Email: Run /inbox when I have energy, archive/todo/skip through them
  4. Tasks: Work from my TickTick “Today” view - everything already has verbs and due dates
  5. Meals: Follow the grocery list, cook the planned meals

The stuff I used to burn executive function on - figuring out what to do, remembering how to do exercises correctly, deciding what to eat, enriching budget transactions, processing email - just happens.

My job is now to:

  1. Show up at the gym
  2. Eat what’s on the list
  3. Work through the tasks Claude helped me prioritize

The meta-work is gone. The actual work remains.

The Technical Stack

For the curious:

Claude Code Setup:

  • Custom slash commands defined in .claude/commands/*.md
  • Each command is a workflow written in markdown that Claude follows
  • MCP (Model Context Protocol) servers for Hevy, Slack, Linear
  • Python scripts for Gmail, TickTick (Claude calls via Bash)

AWS Infrastructure:

  • Lambda (container-based) for background jobs
  • ECR for container images
  • SSM Parameter Store for all credentials (“Claude 1Password”)
  • EventBridge for hourly scheduling
  • Bedrock (Claude Haiku) for text processing in Lambda
  • SNS + ntfy.sh for mobile alerts
  • CloudWatch for logs and metrics
  • Terraform for infrastructure-as-code

Integrations:

  • Gmail API (personal + work accounts)
  • YNAB API (budgeting)
  • TickTick API (tasks)
  • Linear API (issues)
  • GitHub API (PRs)
  • Hevy API (workouts)
  • Slack API (messages)

The slash commands are just markdown files that define a workflow. Claude reads them and follows the steps. It’s programming by writing English that explains what you want.

What This Means

I’m not sure I’m “using AI” anymore in the way most people mean. Claude isn’t a tool I pick up to do a task. It’s the interface through which I interact with my own life systems.

The distinction matters because it changes the frame from “AI can help with X” to “what would it look like if AI handled all the glue work?”

For someone with ADHD, the glue work - context switching, remembering, prioritizing, maintaining systems, generating options, sustaining focus - is often harder than the actual work. The actual work is fine. It’s everything around the work that kills me.

Now I have a system that:

  • Aggregates instead of scatters
  • Enforces methodology instead of hoping I remember
  • Automates the boring parts entirely
  • Guides me through decisions with proposed options
  • Remembers everything I forget

I just have to show up and answer questions.

Getting Started

If you want to build something like this:

  1. Start with one pain point. Mine was the morning chaos. What’s yours?
  2. Write a Claude command for it. Just a markdown file explaining the workflow you wish existed.
  3. Add integrations as needed. MCP servers, Python scripts, API calls.
  4. Notice what else is annoying. Build a command for that too.
  5. Let it compound. Each new piece makes the whole system more useful.

The system grows from there. You’re not building an app. You’re programming a chief of staff in English.

If this resonates and you want to see the actual code/commands, I might open-source the setup. Let me know.