I Stopped Trying to Fix My ADHD. I Built Infrastructure to Replace My Memory Instead.

Most productivity systems are built by people without ADHD. That is why they fail people with ADHD.

They assume discipline. A weekly review. A daily log. A habit of tagging things in the right place at the right time. And maintaining things is precisely what ADHD makes impossible. The system sounds reasonable until the first bad week, and then it collapses, and you blame yourself.

The problem is not you. The problem is any system that requires you to maintain it.

Since December 2025, I have been running my entire business through what I call a Second Brain: a Claude Code-powered work system backed by a database I own, automated so that nothing depends on me remembering to do it. Every part of this system exists because I asked one question: where does my discipline fail, and what can I build there instead?

This is a log of what one day actually looks like. Including the parts where things break.

---

The design principle

The Second Brain is not a collection of AI tools. It is plumbing. And the design principle behind every piece of it is the same: if it requires me to remember, it will fail.

The foundation is a Supabase PostgreSQL database. CRM records, purchase history, tasks, newsletter data, accounting transactions, session notes. Everything writes to it automatically. I do not maintain it. I query it.

On top of that sit MCP servers: API bridges that let Claude Code talk to Gmail, WhatsApp, Stripe, Google Calendar, LinkedIn, and GitHub directly. No browser tabs. No copy-paste. The system reads and writes to external tools the way I would, except it does not get distracted halfway through.

The interface is 60+ custom slash commands. /begin starts a tracked session. /end closes it and opens a pull request. /respond-second-brain drafts outreach. /morning-briefing compiles my day. Each one is a workflow that does not require me to remember the steps.

And underneath all of it: hooks. Shell scripts that fire automatically. Send an email, and a hook logs it to the CRM. Send a WhatsApp, same thing. Try to commit code to main branch, and a hook blocks you with an explanation. I do not have to remember to log anything. The system logs for me.

That is the design. Not "AI makes you productive." Infrastructure replaces the parts of work where your brain fails.

---

6:47am. Before coffee.

The first thing I do every morning is not check email. I run a briefing agent.

/morning-briefing

The agent opens Gmail, WhatsApp, and my CRM simultaneously. It reads my calendar. It scans my prospect table for overdue follow-ups. It pulls my active tasks. Then it writes a plain-language summary: what arrived overnight, what is overdue, what needs a decision today.

Here is why this matters for ADHD specifically: left to my own devices, I will spend 40 minutes in email rabbit holes and miss the one thing that needed to happen by noon. The briefing is not a productivity hack. It is a hard external structure that tells me what the day is before I start reacting to it.

The agent does not motivate me. It reads data and reports facts. On routine days the summary is four lines. On chaotic days it is twelve. Either way, I know where I am before I open anything else.

Design vs discipline: a disciplined person checks their calendar, reviews their CRM, scans their inbox in order. I built a system that does all three in parallel and hands me a summary. Same outcome, no willpower required.

---

9:15am. Creating content.

I have 2,238 newsletter subscribers. I write two or three pieces a week. I also post on LinkedIn. Doing this manually while building a product and managing 85 clients would require more sustained attention than I have.

/content-team

This spawns two agents sequentially: a creator and an evaluator. They do not share context. The creator writes based on a brief I give it. The evaluator scores the output against a rubric without seeing the creative process. No confirmation bias.

The evaluator uses a 100-point framework: hook strength, voice consistency, engagement mechanics, structural integrity. Below 70, the creator revises. Below 70 after two rounds, I read it myself and figure out what is wrong.

This morning I needed a LinkedIn post about AI memory systems. Two-sentence brief. The creator wrote three options. The evaluator scored them 68, 74, and 71. The Educator version got flagged for a weak hook and uniform bullet lengths. The creator rewrote the hook. Score moved to 81. Total time: 22 minutes.

Design vs discipline: a disciplined person would write, edit, rewrite, and resist the urge to publish the first draft. I get bored after the first draft. So I built an external editor that does not get bored.

---

10:30am. Sales pipeline.

I have a prospect named Marcin. He commented on a LinkedIn post three weeks ago, signed up for the newsletter, then went quiet. The CRM shows his last contact date, the channel, and my note: "interested in Kickstart, hesitated on timeline."

/respond-second-brain

The skill reads his full CRM record, scrapes his LinkedIn for recent activity, pulls my previous email thread. It drafts one email and one LinkedIn DM. Both reference the specific context from our previous conversation. Neither says "just checking in."

I change one sentence. I send both.

What happens next requires no action from me. The Gmail hook catches the send and writes a CRM entry: contact date, channel, summary. WhatsApp hook does the same. I do not log outreach manually.

This is what most people miss. Every AI tool can draft an email now. That is not the point. The point is continuity. Next week, when I think about Marcin again, the system already knows what happened this week. I do not have to reconstruct it from memory. I cannot reconstruct it from memory. That is what ADHD means in practice.

The CRM tracks over 200 prospects. I manage all of them this way. No mental overhead for state.

Design vs discipline: a disciplined salesperson updates their CRM after every interaction. I have never once remembered to do that. So I built hooks that do it the moment the interaction happens.

---

11:45am. Building product.

I am adding a new feature: a Kanban-style task view for the ADHD brain playground. Before I touch a single file:

/begin adhd-brain-kanban-playground

This creates a GitHub issue, checks out a new branch, registers the session in my task database, and activates the branch guard hook. The branch guard blocks any code edit or commit on main. If I accidentally try to push to production, the hook fires and stops me.

I spawn an agent team. Planner, builder, reviewer. Non-overlapping files. Two agents editing the same file causes silent overwrites. I learned this by losing work.

Today the reviewer caught a missing .js extension in an API import. On Vercel, this causes a 30-second timeout that kills the request. Caught before I pushed anything.

The agents do not write brilliant code. They handle the mechanical parts, catch the dumb mistakes I make when my focus drifts after hour two, and keep the session structured enough that I can pick it up after an interruption.

4,351 tasks tracked since December 18, 2025. 414 executed autonomously by agents. Real numbers from a real database table.

Design vs discipline: a disciplined developer remembers to create branches, write tests, review their own code before pushing. I forget. So the system refuses to let me skip those steps.

---

1:30pm. When things break.

While building the Second Brain, I am also using it to build the Second Brain. The system is both the product and the workshop.

This morning I noticed the branch guard was firing incorrectly on .json files. I filed a task. The agent team investigated. The builder patched it. The reviewer validated. The PR merged. The hook now works correctly.

That loop happens constantly. I use the system, notice something wrong, fix it with agents, and the system improves. The next session starts from a better baseline.

But here is the honest part.

Last week I tried to automate Zoom registration using browser automation. The UI changed mid-sequence. The automation failed three times at different steps because Zoom's button IDs are not stable. I gave up and registered people manually. The automation is still on my task list.

Two weeks ago, a linter ran on a file I had not touched and removed code used elsewhere. The build broke. I spent 45 minutes finding it. Root cause: misconfigured ESLint rule. I fixed the rule, added a test. But the 45 minutes happened.

And there is the context window problem. When a session runs long and Claude's context fills up, it compacts automatically. Sometimes the compaction drops context I needed. I now write explicit session notes at checkpoints. Not because the system told me to. Because I burned myself twice by not doing it.

These failures are the actual texture of using AI tooling in production. The system makes things faster and more reliable. It does not make things perfect. Design replaces discipline, but it does not replace judgment.

---

2:00pm. Accounting.

My ADHD makes financial tracking painful. Sustained attention to detail with no immediate reward. I put it off. Then I regret it.

The accounting system imports CSV exports from my Polish bank and Revolut. It parses transactions, categorizes them, reconciles against expected income and expenses, and generates a financial health report.

The first time I ran reconciliation, it flagged a mismatch between Stripe revenue and bank deposits. One payment had settled in a different month. The system caught it in seconds. I would have found this at tax filing, if at all.

$48,573 from 82 paying customers in four months. That number came from a SQL query cross-referenced against Stripe records. It is accurate because the system tracks it automatically, not because I am disciplined about bookkeeping.

Design vs discipline: a disciplined person reconciles their accounts monthly. I built a system that does it in one command and flags the discrepancies for me.

---

4:15pm. Support.

A client emails about a questionnaire error. Their answers were not saved.

I query their purchase record. I see the Kickstart package, the webhook, the exact timestamp, the email confirmation, the onboarding stage. I reply with a specific answer. Not "let me look into this." I know what happened. I fix it. Five minutes.

85 clients at my current bandwidth requires this kind of operational leverage. I am one person. I cannot hold all of this in my head. The system holds it.

---

6:30pm. Closing out.

/end

The session closer commits the branch, creates a pull request, updates the task in the database, closes the GitHub issue. I write two sentences of notes.

The data tells me the truth even when I am not paying attention. On weeks where the task count drops, something is wrong with my energy or my focus. The data makes that visible faster than my mood does.

That is what I find most useful about the whole system. Not the automation or the AI drafting. The measurement. The system does not make me productive. It removes the parts of work that my ADHD brain is worst at: sustained tracking, state recall, context reconstruction, consistent follow-through. What remains is judgment, decisions, relationships, creative work.

That is a better division of labor. Not because AI is taking over. Because my brain has specific failure modes, and I built something that compensates for them instead of asking me to overcome them.

---

If you want to know whether your own setup has the same gaps, the health check is free:

claude mcp add second-brain-health -- npx second-brain-health-check

Run check_health. It scores your setup across 45 dimensions and tells you exactly where the gaps are. No pitch. Just the number.

What part of your current workflow requires you to remember to use it? That is exactly where it will break.

PS: I run daily coworking sessions where I do this live. 25 minutes, every weekday. Two times: 10am CET and 10am ET. No slides, no pitch. Just me working with the system on whatever is on my actual to-do list. Join free at iwoszapar.com/coworking