← Back to Autonomy
Copilot Kit · Visual Skills · Demo Reel · By Majid Mazouchi

Visual Skills Showcase
26 skills · 26 live samples

Every section below is a working sample of one skill from the visual-communication kit — built exactly the way the corresponding SKILL.md instructs an agent to build it. View source on any sample to see the pattern.

§ 01

The page system itself

skill: paper-monograph-shell

This page is the sample: hero, mono-numbered TOC, sectioned layout, figure and lab conventions — all from the shell skill. Below, its two core vocabularies: the semantic token palette (same meaning in every figure on a page) and the standard readout chip.

violet · model / reasoning
orange · actions / commands
blue · feedback / context
green · environment / good
red · errors / failures
gold · gates / budgets
Healthy metric
62.3°
Marginal metric
4.8 dB
Failing metric
UNSTABLE
Callout card convention

Five-pixel left border in a semantic color, mono uppercase label, one focused insight. Used for "the core trade", "failure smell", "practical heuristic".

§ 02 · Interactive

Slider-driven live simulation

skill: interactive-parameter-labs

The slider→recompute→redraw pattern: a second-order system simulated by forward Euler on every input event, with threshold-colored readouts. Drag damping from underdamped ringing to overdamped crawl.

Sample lab · second-order step response
Overshoot
Settling ±2%
y(t) for a unit step; dashed line = setpoint. Euler, dt = 5 ms, recomputed live.
§ 03 · Interactive

Algorithm walkthrough with synchronized code

skill: algorithm-step-player

Precomputed immutable snapshots + a pure render function: binary search for 37, with the active window in blue, the probe in orange, eliminated cells faded, and the matching code line highlighted at every step.

Sample player · binary search
Step
0
Comparisons
0
§ 04 · Interactive

Clickable decision flowchart with path walking

skill: interactive-flowcharts

Tap any node for detail. Then press Walk scenario to watch a doom-loop diagnosis traced through the chart — the path-walking move that makes flowcharts presentable instead of static.

Sample flowchart · agent-loop triage
yes no yes no yes no Tests failing? Same error 2×? Lint+types ok? L-02 doom loop → fix error legibility, cap retries Iterate: one hypothesis, re-verify Stop gate ✓ ship
Tap a node, or walk the scenario.
§ 05 · Interactive

Scripted state-machine replay

skill: state-machine-and-sequence-animations

A three-state machine driven by an event script: the current state glows, the firing transition flashes orange, and every event lands in the log — the same transport controls as the step player.

Sample FSM · health-monitor mode logic
start done fault_detected reset IDLE RUN FAULT
§ 06 · Interactive

Hover-highlight, layer toggles, tooltips

skill: svg-interactivity-layer

Hover (or tab to) any block: its connected neighborhood stays lit while everything else dims, and a tooltip follows the pointer. The checkbox toggles a teaching overlay — the fault-injection layer — without reflowing the diagram.

Sample · explorable signal chain
feedback sensor bias / drift injection load fault Sensor Filter Controller Plant
§ 07

Executive dashboard band

skill: exec-dashboard-figures

The 3-seconds-from-2-meters discipline: big numbers with meaning-colored deltas, a percent-true gauge (pure SVG, no JS), a sparkline for shape, and same-scale comparison bars. One accent per band, one takeaway sentence.

Recall scope (proposed)
$2.4M
▼ 38% vs. baseline method
Mean time to diagnosis
42 min
▼ 18% vs. Q1
Escalations to teardown
61
▲ 9% vs. Q1 — see §4 note
Detection coverage at reduced scope
73% target ≥ 70%
12-week trend
Vehicles in recall scope (indexed, baseline = 100)
Baseline VIN-range method · 100
Proposed telemetry-scoped · 62
+ usage-conditioned · 47

Takeaway: telemetry-scoped selection cuts recall scope 38% while holding detection coverage above target.

§ 08 · Interactive

Embedded slide deck

skill: html-slide-mode

Click the deck once, then drive it with ← → or space (buttons work too). The same engine turns a full monograph into a talk: sections become slides, labs present live, and a print stylesheet exports the deck to PDF.

Sample deck · slide engine demo

One artifact, three formats

The same HTML file is a readable document, a keyboard-driven deck, and a printable PDF.

Principle

One idea per slide

  • Cut prose to ≤ 3 lines; the figcaption carries nuance.
  • Every figure or interactive lab gets its own slide.
  • Presenter notes hide behind a flag — projector never sees them.
Why it works for research talks

Slides that are demos

Interactive labs keep working inside slides — tune a parameter live in front of the audience instead of showing a screenshot of a plot.

End

← → to navigate · print to export

~40 lines of vanilla JS. No reveal.js, no build step, works offline.

1 / 4 focus the deck, then ← → / space
§ 09

Animated block diagram

skill: animated-svg-block-diagrams

The original skill of the family: hand-coded inline SVG — boxes on a manual grid, one shared arrowhead marker, two-line labels — with the dash-flow animation that makes the signal visibly travel the loop. Three lines of CSS, frozen automatically under reduced-motion.

command r Σ e = r − y Controller PID · lead-lag · agent effort u Plant the physics output y Sensor measurement (− y)
Fig. 1 — The signature pattern. Semantic colors (orange command, red error, violet controller, green plant, blue feedback), a single marker def reused by every arrow, and dashes that travel in the signal's direction — the whole technique the skill teaches, in one figure.
§ 10 · Interactive

Radial mind map with collapsible branches

skill: mind-maps

One color per branch family, curved spokes, horizontal labels only — and each first-level branch collapses on click (the −/+ glyph), keeping portfolio overviews navigable.

Sample · VHM idea portfolio at a glance — click a colored branch node
multi-agent MCP / A2A guardrails anomaly det. RUL models flux-map NN observers parity eqs Kalman / UKF adaptive DAQ edge triggers compression VHM portfolio Agentic AI ML Model-based Data acquisition
§ 11 · Interactive

Fault tree with cut-set highlighting

skill: interactive-tree-diagrams

FTA conventions in SVG: rects for the top and intermediate events, circles for basic events, the pointed arch for OR and the flat-shouldered dome for AND. The button lights up a minimal cut set — leaves, gate, and connecting path together.

Sample fault tree · torque-accuracy loss
Loss of torque accuracy OR Sensor path fault Model path error OR AND Resolver offset ADC drift Flux map stale Temp est. error
Cut sets not highlighted.
§ 12 · Interactive

Concept graph with typed edges

skill: concept-graphs

Not a tree, not a mind map: arbitrary relations, hand-placed for zero crossings. Solid edges = depends-on, dashed = informs. Hover or tab to any node to isolate its neighborhood.

Sample · how the research threads connect
Telemetry Diagnostics VHM core Prognostics Recall optimization Fleet analytics
solid = depends-on · dashed = informs · hover a node to isolate its neighborhood
§ 13 · Interactive

Live weighted trade study

skill: trade-study-matrices

Raw scores stay visible, weights move on sliders, the ranking recomputes live, and a sensitivity line says whether any single weight could flip the winner — hand a stakeholder the slider instead of an argument.

Sample · ranking three portfolio ideas (scores 1–10, illustrative)
OptionImpactFeasibilityTime-to-demoWeighted
VHM-016 Recall scope opt.965
VHM-007 Adaptive DAQ787
VHM-023 Fleet triage agent698
§ 14 · Interactive

Risk matrix with clickable register chips

skill: risk-and-heat-matrices

Likelihood descending down the side, severity across the bottom, muted band fills with a band letter in every cell (color never carries meaning alone), and risk chips that open their register entry.

Sample · 4×4 risk grid — tap a chip
Tap a risk chip for its register entry.
§ 15

Roadmap with swimlanes, milestones, today-marker

skill: timelines-and-roadmaps

Every x-position derives from one date→pixel function: quarter gridlines, phase bars, gold milestone diamonds, and the dashed red today line that keeps a roadmap honest.

Q3 '26 Q4 '26 Q1 '27 Q2 '27 Research Prototype Pilot Ideation + literature review Adaptive DAQ loop v1 Fleet pilot (50 vehicles) Exec review Gate 2 today
Fig. — Roadmap conventions. Alternating quarter shading, lane separators, labels inside bars only where they fit, milestone labels dodged above the lane.
§ 16

Waterfall: baseline → savings → net

skill: flow-quantity-diagrams

Zero-based scale, floating segments tied together by dashed connectors, colors by meaning (savings green), and the value printed at every segment — geometry for shape, numbers for the data.

$0M $2M $4M $3.9M −$1.0M −$0.5M $2.4M net baseline scope telemetry scoping usage-cond. proposed scope
Fig. — Bridge from baseline to net. The dashed connectors are what make floating segments read as one continuous story; totals reconcile by construction (3.9 − 1.0 − 0.5 = 2.4).
§ 17 · Interactive

Scrollytelling: the figure follows your reading

skill: scrollytelling-figures

Scroll through the three steps below: the sticky figure shifts emphasis as each step enters view (IntersectionObserver, class toggles, reduced-motion fallback shows everything at once).

1 · Gather. The loop begins by assembling context — the task, the relevant files, the last results. Nothing intelligent has happened yet; this stage decides what the model is even able to know.

2 · Act. The model proposes one action — an edit, a command — and the tool layer executes it against the real environment. Ambition is cheap here; the next stage is where it gets priced.

3 · Verify. Tests, types, and lint judge the result without trusting the model, and their verdict flows back into the next iteration's context. This stage — not the model — bounds the loop's reliability.

Gather context assembly Act model + tool call Verify tests · types · lint loop
§ 18 · Interactive

Self-check quiz

skill: self-check-quizzes

Retrieval practice, not grading: the reader commits, the correct answer is always revealed, and the explanation — the real payload — cites where the concept was taught.

Q. In an edit-verify agent loop, raising verifier strength primarily converts…
Right — a stronger verifier catches broken attempts, so they loop again instead of shipping. Failure becomes recoverable; only "budget exhausted" remains, and that's fixable. (This is Lab A of the agentic loop-engineering guide.)
§ 19 · Interactive

Zoom & pan an oversized diagram

skill: zoom-pan-svg

ViewBox manipulation, so strokes and text stay crisp at any zoom: scroll-wheel zooms toward the cursor, drag pans (one finger on touch, pinch to zoom), and Reset always brings you home.

Sample · system landscape — wheel to zoom, drag to pan
VEHICLE CLOUD Inverter ECU iq/id · temps · faults BMS ECU SoC · SoH · cell ΔT Chassis ECU wheel speeds · IMU Gateway CAN/Ethernet bridge Telematics edge triggers · buffer OTA uplink Ingest decode · validate Feature store per-VIN time series VHM models diagnosis · RUL · triage Recall optimizer scope selection · ROI
zoom: 100%
§ 20 · Interactive (loads three.js)

The 3D foundation

skill: threejs-scene-foundation

The one place the kit permits a library — with obligations: load on demand (never block the page on 600 KB of engine), pin the version via import map, vendor the module for offline pages, and fall back gracefully. Press the button to load three.js and bring §§20–25 alive; everything else on this page stays zero-dependency.

Load once, power six scenes
not loaded — requires network once
The foundation demo: a stylized pickup assembled from ~15 primitives — body, greenhouse, bed rails, grille, lights, wheels — on the paper-styled stage. Drag to orbit, scroll to zoom. For a real production model you would load an engineering glTF via the exploded-model-walkthroughs skill.
§ 21 · Interactive (3D)

Data surface with hover readout

skill: interactive-3d-surface-plots

A displaced plane with per-vertex token colormap — the pattern behind efficiency maps and flux-linkage surfaces ψ(id, iq). Move the pointer over the surface to read values off it; drag the ripple slider to watch live regeneration (same buffers, no reallocation).

Sample surface · z = dome + ripple
Hover readout
load 3D, then hover
Colormap: blue→green→gold→red by height, built from the page tokens. Raycaster converts the hover point back to data units.
§ 22 · Interactive (3D)

3D system diagram with flow particles

skill: 3d-system-diagrams

Extruded token-colored blocks on a ground grid, tube connections, and orange particles riding the feedback arc — the 3D dash-flow. (The skill adds crisp HTML labels via CSS2DRenderer; this demo keeps labels in the caption.)

Sample · model → tools → environment, feedback overhead
violet = model · orange = tools · green = environment · blue arc = feedback path, particle direction = data direction
§ 23 · Interactive (3D)

Exploded view on a slider

skill: exploded-model-walkthroughs

No glTF asset needed to explain a machine: primitive mockups (shaft, rotor, stator, housing) with stored rest poses and per-part explode directions. Drag the slider; 0 reassembles exactly.

Sample · schematic e-machine, axial explode
gray shaft · green rotor · blue stator ring · gold housing ring — directions chosen so the explosion reads as disassembly order
§ 24 · Interactive (3D)

Rotating frames & live vectors

skill: physics-and-motion-demos

Clock-driven motion (frame-rate independent by construction): gray stator phase axes at 0°/120°/240°, the orange current space vector, and the blue/green d–q frame rotating with it — the Park transform as an animation.

Sample · space vector and dq frame
gray = stator phase axes (static) · orange = current space vector · blue = d-axis · green = q-axis. Orbit to see it's genuinely planar.
§ 25 · Interactive (3D)

Camera choreography

skill: 3d-camera-choreography

Named poses + one smoothstep tween that moves position and target together. Fly between views; grab the scene mid-flight and the tween yields to you. Bound to slides or scroll steps, this turns any scene into a guided tour.

Sample · three poses over a block-scape
Reduced-motion preference makes every move a jump-cut automatically.
§ 26

Isometric SVG — depth with zero dependencies

skill: isometric-svg-diagrams

The question to ask before reaching for WebGL: does it need to orbit? If not, two lines of projection math give you depth in pure SVG — offline, printable, weightless. Three shaded faces per box, one light direction, painter's draw order.

Apps · agents Platform · VHM services Edge · vehicles triage agent recall optimizer
Fig. — Layered platform, no WebGL. Every vertex computed through sx = X₀ + (x−y)·cos30°·k, sy = Y₀ + (x+y)·½·k − z·k; top faces lightest, right faces darkest, one light direction page-wide; layers drawn bottom-up (painter's order).