First Step

• 

  Logo

• 

  Background Graphic

Inspiration

I did not build First Step from an abstract productivity thesis. I built it because I have ADHD, and I struggle every day to focus while managing an endless stream of thoughts and ideas. Mundane but important tasks are often the first to be starved of attention; I can remember one and lose it again seconds later.

Hyperfocus used to carry me surprisingly far. Now I am a dad, a husband, a dog owner, and a startup founder. I need to remember errands, care for a household and the people in it, and keep a company moving forward—often in the same day. Hyperfocus on one thing no longer solves the problem; it can make everything outside that tunnel disappear. The motivation for First Step is deeply personal: my well-being, and my ability to show up for the life I have built.

I tried task managers, note-taking apps, and productivity systems. Their shared problem was that capturing a thought still created work. They asked me to decide whether it was a task or a note, choose a project, add a priority, estimate it, categorise it, break it down, or return later for a sorting session. The tools still needed my brain to organise the very thoughts my brain was struggling to hold.

LLMs changed what was possible. A carefully bounded agent can accept the thought as it arrives, preserve it, understand it, and do much of that organisational work without demanding more input first. I wanted to make something quieter than another to-do list: an external brain where I can unload what is in my head without self-editing, then receive one appropriately small next step when I have the capacity to act.

The product began as a form of self-rescue, then I tested the idea against ADHD research and five personas spanning creators, founders, researchers, business owners, and operations workers. First Step is a productivity and well-being support tool, not a diagnostic or clinical product.

What it does

Capture now; decide later. A single quick-capture surface accepts notes, standalone links, pasted email, and supported files. The person does not need to choose a category, project, priority, or estimate first. The raw capture is safely stored before AI processing, so the thought can leave their head without being lost or prematurely reshaped.

Turn ambiguity into a next action. GPT-5.6 uses Structured Outputs to turn raw captures into strictly validated tasks or memories. For a project, First Step can reveal one visible 5–15 minute action while keeping later steps out of sight until they are useful. If the action is still too big, the user can make it smaller without losing the earlier version.

Meet the person where they are. Before focus, the user states available time, current capacity, and preferred mode. Deterministic ranking combines those explicit choices with deadline and momentum signals to recommend one task that fits now. The optional adaptive layer is deliberately bounded: it learns only from meaningful starts and “show another” choices, and never overrides what the person explicitly said about time, capacity, or mode.

Make returning easy. A focus session can be open-ended, five minutes, or fifteen minutes. The user can pause with a factual Return Point, set work aside without calling it failure, and ask for one grounded next move only when they choose “I'm stuck.”

Remember without re-searching everything. The Brain holds captures, source evidence, history, and open loops. It uses PostgreSQL full-text search plus pgvector hybrid retrieval so answers are grounded in the user's own material rather than a generic productivity script.

How I built it

First Step is a Next.js/TypeScript application deployed on Vercel. Supabase provides invite-only authentication, durable data, private file storage, PostgreSQL search, and pgvector. Every brain, capture, task, retrieval result, and private file is owner-scoped; browser code never receives privileged service credentials or the OpenAI key.

I built it with Codex using GPT-5.6 in iterative, testable delivery slices. I translated my lived experience, research findings, and persona needs into focused build briefs, then used Codex to accelerate the UI, server routes, database boundary, validation, and verification work.

Inside the product, GPT-5.6 Luna handles structured extraction from raw captures, while GPT-5.6 Terra handles ambiguity, grounded recall, and optional stuck-help proposals. This combination matters: the models are capable enough to understand messy, everyday input while remaining affordable enough for a capture-everything workflow. A user should not have to decide that a thought is “important enough for AI” before unloading it. Structured outputs and strict validation keep the intelligence useful and predictable; the agent proposes, while the person remains in control.

Challenges I ran into

The hard part was not adding more productivity features. It was removing the hidden labour those features usually create. A conventional task manager exposes a backlog, priorities, dates, labels, and settings—and often expects the user to maintain all of them. For a person already overloaded, the organisation system can become another unfinished responsibility.

So I designed around progressive disclosure: one clear current action in Focus, with the full Brain available only when the person asks for it. The agent needed enough freedom to sort, clarify, and break work down, but clear limits around authority, cost, and privacy. Raw capture is retained before AI processing, hidden steps are not returned to the browser, and the AI cannot silently mutate tasks, overrule stated capacity, or infer emotion.

What I learned

For ADHD-supportive software, fewer decisions can be a real feature. The most helpful AI experience is sometimes simply permission to dump the unfiltered thought into one box and move on, trusting that it will still be there and can be made useful later.

A useful agent should reduce the cost of remembering, sorting, starting, and returning—not seize authority from the person using it. The strongest product decision was to make “capture now, start when ready” the default.

What's next

I plan to expand the capture channels and continue testing recovery-friendly review patterns, while preserving the quiet Focus surface and the user's final say over recommendations.

Built With

• chatgpt
• codex
• gpt5.6
• next.js
• pgvector
• postgresql
• supabase