Opportunity Scout

May 2026WebPrototype

A made-up test user: a student looking for scholarships and internships.

Opportunity Scout is an AI research team for one person: it learns who you are, then finds the grants, programs, competitions and contacts you would otherwise miss.

It turns your context into a profile, picks sources worth watching, scans them on a schedule, and only lets through the findings that clear its filters, each with why it matters and what to do.

I specced it as 24 systems and had 24 AI workers build them in parallel, and a week later I started Crawft the same way.

systems, one AI worker each: 24
rounds of testing: 13
made-up test users: 12

The best opportunities are the easiest to miss

The things that can change a year, like a grant, a fellowship, a pitch competition or the right introduction, are scattered across directories, newsletters, X threads and program pages. Most people hear about them after the deadline, and the people who don't usually have someone scanning for them. I wanted that someone for everyone: a founder looking for funding, a student hunting scholarships, a career switcher looking for a way in.

Opportunity Scout is that research team. You tell it about yourself once, and it does the watching. What reaches you is one inbox of findings sorted into Act soon, High leverage, Contacts and Later, each with a relevance score, the evidence behind it and a single next action.

Scout runs as a loop that learns

Behind the inbox, Scout works as a continuous loop. Your context becomes a structured profile, and the profile decides which sources are worth watching. A research swarm scans those sources on a schedule and stores everything it finds. Each raw finding is turned into a clean candidate, filtered against your profile, and then checked to make sure it's real, current and actionable before it becomes a recommendation. Urgent ones arrive as alerts; the rest wait in the feed. Everything you do with a finding, saving it, finishing it, dismissing it, feeds back into the profile and the sources, so the next scan aims better.

How Scout finds what matters to you

Using it

Scout is built so you give it context once and keep steering it after that.

Step 1: Give it context once
You describe your goals, projects and constraints. Scout matches them against live sources, from funding directories to communities, filters every find for fit, urgency and evidence, and turns the few that pass into alerts that say what to do next.
Step 2: Steer where it looks
Every source gets a grade based on how much it actually yields. You can add places you trust and remove ones that stop helping; a noisy source like general startup news drops to a cooldown on its own.
Step 3: Trace any recommendation back
The Systems page shows every stage of the engine with an honest state: Ready, Needs Review, Needs Setup or Not Started. When a recommendation looks wrong, you can follow the run back to the stage that produced it, and Ask Scout why explains any card in plain words.

Opportunity Scout's 'How it works' section: four cards for Your context, Live sources, Scout filters and Useful alert, joined by a line, above 'Only the few useful finds reach you.'

The Systems page grid titled 'Every stage has an honest state': 24 cards from Raw Context Capture to Observability and Evaluation, each tagged Ready, Needs Setup, Needs Review or Not Started.

I wrote the product as 24 systems before any code

Scout has a lot of moving parts, and I knew a single AI session would lose the thread long before it finished. So I started with a written spec and broke the product into 24 numbered systems, each with its own job, inputs and outputs: raw context capture, profile building, source discovery, the research swarm, relevance filtering, verification, notifications, feedback learning, billing, privacy and more. A shared contract pinned down the routes and data shapes every system had to use, so 24 pieces built separately would still fit together.

A grid of Opportunity Scout's 24 numbered systems, each with a one-line job, from 1 Raw Context Capture and 10 Research Orchestration to 24 Observability and Evaluation.

Then 24 AI workers built it at once

On May 6 I launched 24 AI coding workers in parallel, one per system, each with its own brief and the shared contract. An orchestrator reviewed what each one returned, ran the tests and opened the next wave of fixes. Over the next week it ran 13 rounds of testing against a dozen made-up test users, from a student hunting scholarships to an indie game studio looking for funding, with an overnight watchdog checking progress every hour. The later rounds rendered every test user's pages in a real browser, so I could see whether the right finding reached the right person.

How Scout was built

Every worker built against the same contract, so 24 pieces built separately still fit together.

Six of the twelve test users, each with a different top finding.

1 of 1

The method went straight into Crawft

Scout proved something bigger than Scout: one person with a clear spec can run two dozen AI workers on one product at once and still end up with a coherent system. On May 14 I paused Scout, and on May 21 I opened the conversation that became Crawft. Crawft was built the same way, a written spec split into systems and many AI workers building them in parallel, at a much larger scale.

Opportunity Scout's landing page: the serif headline 'Find opportunities you would have missed' beside a 3D laptop on a cream background.

The best opportunities are the easiest to miss

Scout runs as a loop that learns

I wrote the product as 24 systems before any code

Then 24 AI workers built it at once

The method went straight into Crawft

The best opportunities are the easiest to miss

Scout runs as a loop that learns

Using it

Step 1: Give it context once

Step 2: Steer where it looks

Step 3: Trace any recommendation back

I wrote the product as 24 systems before any code

Then 24 AI workers built it at once

The method went straight into Crawft

The best opportunities are the easiest to miss

Scout runs as a loop that learns

Using it

Step 1: Give it context once

Step 2: Steer where it looks

Step 3: Trace any recommendation back

I wrote the product as 24 systems before any code

Then 24 AI workers built it at once

The method went straight into Crawft