Give the human a felt, trustworthy sense of what's really there — its shape, what carries weight, where it's drifted or unwired, and where reality has actually bitten — so they can trust the AI and steer it, without dropping back into re-checking everything.
Built with Robert via a recognition loop · 2026-06-18 · the box Claude Design works from · jobs + judgment, NOT mechanisms
01
Who it's for + the two moments
Design for a real person in a real moment, not an abstract user. (Generalizes to any maintainer later — not now.)
You — the steerer — as the human half of the human+AI partnership: the one who's handed execution to the AI and needs to see just enough to trust it and steer.
Moment 1
Returning → the reload
"I've been away — catch me up."
Moment 2
Steering → the topography
"Where do I push?"
02
The core job: felt sight (not a dashboard)
A gestalt you grok at a glance — not data to parse. Feel the system's shape and relationships; see the gaps; feel the space between where it is and where your intuition wants it to go — so you steer the AI to close it. Topography is the heart; reload the lighter companion. Nothing shown for its own sake.
03
What it renders: the ambient / overlay principle
Ambient (always on) = the felt sense of what's real. Structure (relationships) · weight (regions sized by load-bearing-ness) · currency (proven-current vs drifted) · orphans (gaps) · the confident-wrong heatmap (where reality has actually bitten — the truest weakness signal; accumulates over use).
Overlays (on demand — ONE per question, never all at once) = answers to a question. First overlay: blast-radius (the real downstream cone) — on-change (the guardrail: before approving a change, see what it touches) and on-focus (lighter everyday probe).
No computed "strength" score — "needs strengthening" is felt from the real signals, never faked as a number. Net: a felt reality you can always read, with probes you summon — the opposite of a dashboard.
04
The gap-to-ideal
Render the is so richly the human feels the gap themselves; surface machine-spottable candidate gaps (orphan / drift / heatmap) they recognize & steer; NEVER fake knowing the ideal (that's Plato building a new cave). Machine owns the is; the human owns the ought. (Marking intent onto the map = a later layer — the thinking-legibility sibling.)
05
Form factor (hybrid — medium follows job)
Reload
A generated card, refreshed on push
CI-on-commit; current as of last push. A thing you read.
Topography
A live, interactive canvas
Overview → zoom → drill; toggle the one overlay; focus a node for blast-radius. Felt sight + probes can't come from a static image.
Both regenerate from the engine's real output — no LLM in the render path — so both always show what's real.
06
Non-negotiables
N-1Render its own blind spots
Show uncovered/unproven as such; never assert a completeness it doesn't have. (Cardinal rule.)
N-2Currency ≠ correctness
Every surface labels which it is — "current at HEAD" is not "correct."
N-3Cheap + always-current
Deterministic regenerate from real state; no LLM in the hot path.
N-4Felt, not parsed
Grokkable whole at a glance (territory metaphor); depth on drill, never a wall.
07
The kill-list — what it must NOT be
A hairball
Everything shown at once.
A dashboard / decoration
Metrics for their own sake.
A map that claims to be the territory
Hides where it can't see.
Cry-wolf
Flags noise as signal.
A second home for a fact
One truth, one place.
08
How we'll know it's good
✓Checkable
Renders every state · one home per fact · the walk ends at an act · no dead ends · shows its own blind spots · regenerates cheaply.
★Taste — stays-human (your call)
Does it produce a genuine felt sense of the system? Does the topography actually let you steer? Does it feel like reality, not a chart?
09
The walk (gate before acceptance)
Walk it as the user, in each moment (returning / steering), against this brief — before any review. A surface nobody walked is not done. (Experience-First.)
10
The cheapest first cut
1 ★
One screen, on real data, walked & reacted to — before building the whole.
Strong candidate: the topography of Plato itself — real nodes/edges/orphans from the engine we just built (10 nodes / 12 edges / 1 orphan) — rendered just enough to test the felt sight, then grow interactivity + overlays. The governing question: the cheapest thing that tells us the direction is right.