The System I Finally Built

The problem

For most of the decade I ran the engine well and neglected the rudder. I had the workflow. I kept the lists, worked next actions, did weekly reviews of a kind. What I rarely did honestly was climb back up to purpose and direction — and the tools didn’t help. Each one made the engine cheap and the rudder expensive.

A GTD practice built this way doesn’t collapse in weeks. It drifts in months. You keep capturing, reviewing, closing actions, and somewhere around month four the list grows faster than you close it, the weight is heavier than it was, and the things you finish no longer obviously ladder up to anything you consciously chose. You were productive. You weren’t headed anywhere. And because the drift is slow, you blame yourself rather than the system.

Why previous attempts failed

Every tool I tried across ten years — Evernote, OmniFocus, Things, Todoist, Notion, Roam, custom spreadsheets — could run the lists. None made the strategic review unavoidable. The runway was one tap away; the horizon was a separate decision, a separate interface, a separate hour I could always defer. Over months, the runway won by default.

The fix wasn’t a better app. The fix was a system designed so the rudder couldn’t be skipped. No tool on the market is built that way, and building it myself had always been out of reach.

What changed

Agentic AI changed the economics of custom software for non-developers. With AI as my execution layer, I built a production system myself — directing the architecture, methodology, security, documentation, UI/UX, and quality bar, while AI wrote and shipped the code under my review. A workflow I had lived with for years could now be made robust and frictionless, on my own infrastructure, in weeks rather than the years it would have taken to hire it done.

For the first time in a decade, the gap between “what I need” and “what I can have” closed.

What I built

A personal operating system that runs on my own server. It takes my workflow and removes the friction that used to erode it. In plain terms, it lets me:

• Capture anything, instantly, from anywhere. A thought on my phone becomes a filed, classified item in one line — no formatting, no decisions at the moment of capture.
• See what I’m waiting on, and act on it. Items where I’m blocked on someone else surface with clear nudge triggers, so nothing goes quiet for weeks without me noticing.
• Know what my focus should be. The dashboard shows what’s next, what’s active, and what matters — filtered by context, energy, and the higher-level direction I’ve already set for myself.
• Review at altitude, every week, without fail. The weekly review starts from my purpose and goals, not from the task list. The rudder is the first thing I see, not the last thing I get to.

The system keeps personal and professional work in strictly separated workspaces — same discipline, same dashboard, different contexts, different people, different data. The AI’s awareness switches with the workspace; a question asked in the work context never touches the personal one.

I built every line of this myself, with AI as the implementer. I own the architecture, the methodology, the security posture, the documentation set, the quality bar, and the UI/UX decisions. The craft is mine; the typing is AI’s.

The architecture, briefly

For readers who care about the technical shape of the solution:

Three-tier AI orchestration.

A conversational partner for architecture and methodology decisions, an agentic coder that implements and commits under versioned, scope-bounded instructions, and a runtime co-pilot embedded in the dashboard and aware of its current state.

Production infrastructure.

AWS EC2, NGINX, systemd services, HTTPS throughout, an authentication gateway on a separate subdomain, and secrets in AWS Parameter Store — never in code.

Data and capture.

Markdown as the canonical, portable, version-controllable data format; Telegram as the secure-webhook capture surface; a custom Flask backend and single-page dashboard as the working surface.

Security and resilience.

Auth gateway in front of the app, daily S3 snapshots, and two independent version-controlled repositories — one for code, one for data — with deploy keys and automated nightly commits.

Versioned methodology.

A principles document, an API contract, an infrastructure document, and a changelog — every architectural change versioned, justified, and traceable back to a stated principle.

Quality and UX.

One consistent pattern library across all seven working zones, the same visual grammar for edits, completions, and transitions, and documentation and inline help treated as first-class deliverables.

None of this is exotic. What’s new is that one person, directing AI, can stand up infrastructure like this in weeks — and hold it to a professional standard while doing so.

Results

After running the system for several months:

• The weekly review is something I do, not something I avoid. Because the horizons frame it, the review climbs back to direction every time — not occasionally, not when I feel like it.
• The list no longer outgrows me. Volume is stable because capture is frictionless and processing is disciplined.
• Direction is visible. I can tell, in any given week, whether the work is laddering up to what I said I wanted — and when it isn’t, I can see why.
• The two-workspace separation holds cleanly under the load of a new senior role.
• Nothing has broken. Production-grade infrastructure, three layers of backup, no drama.

The transferable lessons

I set out to build my GTD system. I did not set out to build a transformation credential. But running the project as an architect with AI as my execution layer produced five convictions directly relevant to the work I do professionally.

Activity is not direction, and most operating models conflate the two. Transformations fail the way my personal system used to fail — not in collapse, but in drift. Programs continue, ceremonies continue, metrics continue, and somewhere around month six the work has stopped laddering up to the intent that started it. The cause is the same: the review never climbs back to the level where direction is actually set. Designing your operating rhythm so the strategic review is unavoidable — not optional — is a transformation lever, not a productivity tip.

The bottleneck is never the model. It is the orchestration. The same AI plays different roles in my stack depending on where it sits. The useful question is never “which model should we use.” It is “where in the workflow does intelligence need to sit, and what does each tier need to know.”

Surgical beats general. Narrow, specific, scope-bounded instructions produce reliable results. An AI asked to “help the team be more productive” will fail; one asked to extract specific artifacts in a specific format will succeed. Specificity is the transformation lever.

Methodology beats technology. If your method is sound, AI compounds it. If it isn’t, AI accelerates the dysfunction. The most common failure mode in enterprise AI is bolting clever tools onto broken processes.

Governance emerges from lived separation, not from policy documents. I enforced workspace separation before I had a commercial reason to. By the time I needed it, the pattern was already working. The transferable model — AI inside the perimeter extracts and sanitizes, only the sanitized artifact crosses the boundary — is exactly what most organizations will spend the next two years trying to articulate.

Why this matters

I lead complex transformations. They fail the way my personal system used to fail — slowly, invisibly, because the review never reaches the altitude at which direction is actually set.

I have been faithful to a method for a decade, felt its failure modes honestly, and built — with AI as execution layer, and with my own hands on the architecture, security, documentation, and quality — a system that finally makes the rudder unavoidable. The difference between productive drift and genuine progress is how often and how honestly you return to the level where direction is set.

That is, when I think about it, exactly what a good transformation looks like.

— Aaron