Scout · a systems design story

You can't ask for what you can't name.

How an always-on system for surfacing unknown unknowns was designed — the reasoning, the failure modes it routes around, and the architecture that came out the other side.

↓ the territory below
The problem

The nagging sense that a better option exists — somewhere you've never looked.

Every substantial piece of work begins inside a vocabulary. If you know the words — the terms of art, the named methods, the canonical tools — you can search, compare, and choose. If you don't, you can't even form the question. The work still happens; it just happens inside a smaller universe than the real one, and you never find out what you didn't consider.

This is the practical meaning of the famous quadrant. Known knowns are your craft. Known unknowns become to-do items. But the fourth quadrant is invisible by definition — and it's where the best option often lives.

The design target: a system that converts the fourth quadrant into the second — on its own initiative.

The founding insight: vocabulary acquisition is the fastest converter of unknown unknowns into known unknowns. Learn what things are called, and the territory becomes searchable.

A concrete example that drove the whole design: wanting a beautiful landing page but not knowing that "scroll-driven" and "scroll-triggered" are different things, that a magnetic button has a name, or that libraries of every named animation pattern already exist. The gap wasn't skill. It was nouns.

The shape of the answer

No single primitive could do this job.

The request was for something autonomous, proactive, and woven through every workflow. But each available primitive fails alone: a skill is smart and deep, yet passive — it waits to be called. A hook is always-on, yet must run in milliseconds, so it can't think. A knowledge base remembers but never initiates. A dashboard surfaces but never investigates.

So the architecture is a division of labor — four organs, each doing the one thing its primitive is actually good at:

The brain

A skill that maps territory

The full methodology — scoping, sweeping, compiling, routing — loaded only when needed. Smart, slow, invoked.

The radar

A hook that never sleeps

Fifty milliseconds of regex on every prompt, watching for the signature of new territory. Dumb, instant, always-on.

The memory

A knowledge base that compounds

Every mapped domain becomes a permanent brief — indexed, cross-linked, never re-derived from scratch.

The surface

A command center that proposes

Top gaps land in an inbox a human reviews. The system suggests; it never decides what gets pursued.

The six decisions

Every piece of the architecture exists because something simpler fails.

01

Ship a map, not an answer.

The obvious design — "research the topic and tell me what to do" — quietly recreates the original problem: you get one answer, chosen from a universe you still can't see. Scout's deliverable is deliberately not an answer. It's a domain brief: the vocabulary, the taxonomy of approaches, the canonical resources, the tribal knowledge, and — most importantly — the questions you didn't know to ask.

Naming precedes choosing. Once the territory has names, every existing tool — research agents, tutors, design systems — suddenly works better, because now you can tell them precisely what you want.

Design ruleScout names things; other tools resolve them. The brief ends with a routing table, not a recommendation.
02

A dumb tripwire with a smart filter.

"Always-on and proactive" is where systems like this usually die. A prior advisor in this same environment ran a sophisticated check on every response — and earned a documented verdict of near-zero hit rate. The lesson wasn't that proactivity fails; it's that expensive proactivity fails. Anything that runs constantly must cost almost nothing and be wrong gracefully.

So detection is split in two. The hook is deliberately stupid: pure regex hunting for uncertainty phrases — "I don't even know what's possible" — and new-endeavor verbs, hard-capped at one fire per session, three per day. No model call, no network, under fifty milliseconds. When it trips, the session model — which has full context, memory, and judgment — decides whether the nudge deserves a single offered line or total silence.

Design ruleCheap trigger, smart filter. Intelligence belongs where context lives — never in the hot path.
03

Blind lenses find what one searcher can't.

One researcher finds what it searches for — which is precisely the trap when the goal is finding what nobody thought to search for. Scout's sweep runs parallel agents that cannot see each other, each owning a different way of seeing: the Lexicon hunts terms of art and taxonomy, the Toolsmith hunts the canonical stack and the curated indexes, the Elder hunts tribal knowledge — what ten-year practitioners know that no documentation says.

Diversity catches what redundancy can't. Three identical researchers converge on the same popular sources; three blind lenses triangulate territory none would find alone. Each prompt carries elicitation questions tuned to break past the obvious: "What did this field call this concept ten years ago?" "What do beginners reliably get wrong in their first month?"

Design ruleFor unknown-unknown discovery, vary the angle of search, not the volume of it.
04

The maker never grades its own work.

A brief that feels comprehensive is worse than no brief — it closes the very curiosity it was meant to open. And systems confidently approve their own mediocre output; that's a structural failure, not a prompting one. So every brief faces a fresh-context verifier: a separate agent that never saw how the work was made, grading against a frozen rubric of binary, countable criteria. Twenty-five vocabulary terms or fail. A curated index and a practitioner community with working URLs, or fail. Questions that each name the concept unlocking them, or fail.

The verifier opens the files, counts the items, and fetches sample URLs itself. Architecture enforces honesty that intention can't.

Design ruleSeparate the doer from the judge. Make every criterion observable, so the grade is a measurement — not an opinion.
05

Knowledge has to land somewhere.

Research that lives in a chat transcript evaporates when the session ends. Scout's output protocol is therefore as designed as its research protocol: raw lens outputs are deposited untouched, the brief is compiled into a permanent indexed wiki, a ledger records every mapped domain, and re-encountering a known domain triggers a delta pass — never a duplicate, never a re-derivation.

Then the system surfaces without seizing control. The top three gaps — never more — are proposed to a command-center inbox where a human promotes or dismisses them. The radar reads the ledger too, which closes the loop: the more territory mapped, the quieter the system gets.

Design rulePropose, don't promote. Write once, compound forever, and let coverage earn silence.
06

Step zero, not another silo.

The graveyard of personal tooling is full of capable systems nobody's workflow ever passes through. Scout was designed as a layer, not a destination: the session-orientation ritual checks the ledger before proposing a plan; the planning and brainstorming skills consume briefs as context; and every gap a brief surfaces routes onward to a tool that already exists — a tutor skill to understand a concept, a grilling skill to stress-test it, live community research to hear practitioners, design skills to apply it.

It begins things and feeds things. It refuses to become a place where work accumulates.

Design ruleIntegrate at the entry points of existing workflows. A step-zero system should make every other tool smarter — and claim none of their jobs.
The architecture

From a half-formed prompt to compounded knowledge.

The full pipeline, as built. Steps one and two run on every prompt, forever. Everything after runs only when territory is genuinely new. Watch the route draw itself as you descend.

01 Radar Every prompt passes a regex tripwire — uncertainty phrasing, new-endeavor verbs. <50ms 1/session 3/day
02 Judgment The session model checks the ledger and the vaults. Covered or routine → total silence. New → one offered line.
03 Fingerprint The task is decomposed into the two-to-five fields it actually touches — "beautiful landing page" becomes motion design, typography, layout systems.
04 Sweep Parallel blind lenses, each owning one way of seeing, each web-grounded, each ignorant of the others.
Lexicon & Map

Terms of art, named techniques, the taxonomy and decision axes of the field.

Toolsmith

The canonical stack, curated indexes, communities, reference implementations.

Elder

Tribal knowledge: heuristics, traps, taste markers, what experts argue about.
05 Gap diff Sweep findings versus what the vaults already hold — already-known separated from net-new, ranked by leverage on the task at hand.
06 Brief Ten sections compiled and filed: vocabulary, map, resources, tribal knowledge, options, the questions you didn't know to ask, routed next actions.
07 Verify A fresh-context agent grades the brief against a frozen rubric — counting, opening, fetching. Required failures send it back.
08 Compound Ledger updated, top gaps proposed to the command-center inbox, each routed to the tool that resolves it. The radar grows quieter as the map grows larger.
First contact

The maiden run paid for the whole system.

The first scouted domain was the one that motivated the design: web animation and motion design — a field the user could see the results of but couldn't name anything in. Three blind lenses, one afternoon. What came back:

The industry-standard toolkit had quietly become free.

GSAP and every formerly-premium plugin — the engines behind award-winning sites — went free for commercial use in 2025. Unknowable without knowing what GSAP was.

The wished-for "library of every amazing animation" already existed.

Curated, named-pattern collections mean a design brief can now say "magnetic button, split-text hero, scrubbed pin section" — and be understood precisely.

The #1 tell of machine-generated motion is uniform timing.

Everything fading in with identical duration and easing. The fix — varied timings, ease-out entrances, reduced-motion support — became review criteria the same day.

VERDICT: PASS 14 of 14 rubric criteria — graded by a fresh-context verifier that counted the terms, opened the files, and fetched the URLs itself.

And the loop closed immediately: the page you are reading was designed under that brief's rules. No two entrances here share a duration. Everything eases out. Motion reduces gracefully if you've asked it to. The route through the architecture literally drew itself as you scrolled.