Human-in-the-loop web studio to tune AI-generated output by eye. Stand up a local interactive studio (sliders, pickers, drag handles) or an inline edit/highlight/comment annotation studio for prose & media, instead of guessing values or shipping a static comparison grid.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Human-in-the-loop web studio to tune AI-generated output by eye. Stand up a local interactive studio (sliders, pickers, drag handles) or an inline edit/highlight/comment annotation studio for prose & media, instead of guessing values or shipping a static comparison grid.
Use when the user says "lookdev", or asks to tune / dial in / iterate on the look of something, compare variations by feel, or review / edit / annotate a blog post, doc, copy, or media set. Use whenever "show me, I'll pick" beats asking the user to specify a number, and whenever you'd otherwise hand back a static grid or a wall of prose for review.
When the user says "lookdev" — or any of: tune, dial in, iterate on the look of, compare variations of, let me adjust, let me edit/annotate/mark up, review this post/doc/copy — they mean build an interactive in-browser tool the user directly manipulates. Not a static grid of N variations. Not a Q&A where they specify numbers. Not a wall of prose they're asked to read and reply to in chat. A real-time studio where they act on the artifact and the change is captured.
Two studio shapes — pick by what's being tuned:
Visual-parameter lookdev — the artifact's look is set by numbers/choices (color, type, layout, image treatment, animation, 3D). Controls = sliders, pickers, drag handles. This is the bulk of this skill (below).
Text & media lookdev — the artifact is a document, blog post, copy, or media set and the user is editing/curating it: rewriting sentences, cutting boring paragraphs, highlighting, leaving margin comments, flagging "diagram goes here" / "wrong image, replace." Controls = direct inline editing + selection highlight + anchored comments + media annotation. See the dedicated section below. A blog post / doc / script review IS this mode — never hand back a long markdown file and ask the user to react in chat. Stand up the annotation studio.
What it covers
Any visual decision the user picks by feel, not by spec. Expand this list as needed:
Anything else where "show me, I'll pick" beats "ask me to specify a number"
Controls must stay reachable while inspecting
If the studio shows a list, grid, or scroll-long set of variations, controls must be visible from every scroll position. The user has to be able to drag a slider while looking at row 14, not scroll back to the top each time.
Two approaches, pick by layout:
Sticky bar (position: sticky; top: 0) at the top of the scroll container. Keep the bar visually distinct — paper background + blur backdrop + bottom border — so it doesn't muddy the specimens scrolling behind it. Sticky pins relative to the nearest scrolling ancestor with a defined boundary; if you nest it inside a sized parent (a <header> with margin-bottom, a <div> with a fixed height), it stops sticking at that parent's bottom edge. Lift it to be a direct child of <body> (or the page-wrap) so stickiness spans the whole page.
Floating overlay (position: fixed) for hotkey-toggled controls — e.g. press d to reveal. The portfolio's .debug-ctl pattern is this: pinned top-left, transparent until summoned. Use when the controls shouldn't occupy permanent screen real estate (final viewers shouldn't see them; the author can summon on demand).
Anti-pattern: a top-of-page control panel that the user scrolls past and never sees again. They will tune blindly, give up, or guess. Either keep the controls in view or duplicate a compact control bar next to each variation row.
Text & media lookdev — direct edit, highlight, comment, annotate
When the artifact is a blog post, doc, copy deck, script, or media set, the user is not turning knobs — they're marking up the work the way an editor marks a manuscript. The studio renders the real artifact WYSIWYG (the actual rendered blog with its real components/media, not a raw-markdown textarea) and lets the user act on it directly. Building this for a doc review is mandatory: do not paste a long file into chat and ask "what do you think?" — that's the boring wall of text the user is rejecting. Stand up the annotation studio and let them edit in place.
The four affordances (build all that apply)
Direct inline editing. Every text block is editable in place — click a paragraph/heading and type. Use contentEditable per block (or click-to-swap-to-<textarea>), each block carrying a stable data-block-id that maps back to a source location (markdown/MDX line range, JSX node, or content key). Capture the edited text per block; the agent applies the diff to source. Don't make them retype in a separate field — they edit the rendered sentence.
Selection highlight. Select text → toolbar (or hotkey) applies a colored highlight (<mark>). Multiple colors = a legend the user defines (e.g. yellow "cut this", green "love it", red "wrong/fact-check"). Each highlight stores {blockId, startOffset, endOffset, color, optional note}.
Anchored comments / margin notes. Select text or click a media region → attach a comment shown in a margin rail (pin in the gutter, expand on hover/click) or as a numbered superscript. Comment = {anchor, text} where anchor is a block+range or a media region. This is how the user says "diagram goes here", "too long, cut to two sentences", "needs a real screenshot".
Media annotation. For images/figures: draw a box / drop a pin / arrow on the image and attach a note ({mediaId, x, y, w, h, note}); plus a per-media flag menu — "replace", "wrong model", "regenerate", "missing — generate one here". Placeholders ("DIAGRAM HERE", "MEDIA?") render as visible drop-zones the user clicks to specify what they want, directly addressing "where are the diagrams / where is the media."
Round-trip is MANDATORY (same rule as the settings JSON)
The studio is worthless if the agent can't read the markup back out. Every edit, highlight, comment, and media-flag must export as one machine-readable patch with a single Copy button (and persist to localStorage/URL so a refresh doesn't lose work — this is human-labeled data; see human-labeled-data-rule). Shape:
{"edits":[{"blockId":"p-12","text":"new rewritten text"}],"highlights":[{"blockId":"p-3","range":[40,88],"color":"cut","note":"boring, drop"}],"comments":[{"anchor":"p-7","text":"diagram goes here — flow of the save loop"}],"media":[{"mediaId":"fig-2","flag":"replace","note":"use a real screenshot, not ASCII"}]}
The agent ingests this and bakes: applies the inline edits to the source file, acts on every comment/flag, swaps/generates the flagged media, resolves the highlights (cut the "cut" spans, etc.). Then re-serve the updated artifact for another pass. No markup may exist that isn't in the export blob — otherwise you're back to the user narrating changes by hand.
Mechanics
Render the real thing. MDX/React blog → mount the actual components; static page → render the real HTML/CSS. WYSIWYG per Architecture #5. An annotation layer over a fake-looking preview lies about the result.
Selection → offsets. Use the Selection/Range API; store character offsets relative to the block's text content (not DOM node paths, which break on re-render). Re-apply highlights/comments on load by walking each block's text to the stored offsets.
Editing toolbar floats with the selection (a small popover at the selection rect) or a sticky top bar — controls stay reachable (see section above). Hotkeys: highlight on a key (e.g. h), comment on c.
Keep edit/annotate modes distinct so a stray click doesn't garble text while they meant to highlight — a mode toggle (Edit · Highlight · Comment) or modifier key.
Everything else — serve locally on a free port, verify headless, tear down after baking — is identical to the visual-parameter workflow below.
Control patterns
Pick controls by what the decision actually is.
Decision type
Control
Continuous value (intensity, size, opacity, k)
<input type=range>paired with an editable <input type=number> (not a static label) — drag OR click-and-type; they two-way sync
Discrete choice (mode, blend, easing kind)
segmented buttons or radio chips
Color
<input type=color> swatches; pre-extract dominant palette with coverage % when relevant
Position / size on a canvas
drag the element itself — handles, not numeric inputs
Crop region
draggable rectangle + aspect-lock toggle
Multiple discrete states
render each in a labeled card on one page
Font choice
searchable picker + editable sample-text input
Spatial rule: if the user could point at the thing and drag it, that is the control. Don't add an x: slider when a drag handle is the obvious affordance.
Gesture capture — never make the gesturing hand leave the gesture. When a control toggles a live mouse action the user is performing — recording a cursor path, scrubbing, freehand-drawing, demonstrating a motion — the start/stop must not be a button they have to click. Clicking it drags the mouse off the path, pollutes the start/end of the very motion being captured, and forces a round-trip back to where they were. Bind start/stop to the keyboard (spacebar by default) — keydown on Space, e.preventDefault() to kill page scroll, toggle the same handler the button would. Keep the button too (discoverability), but the hotkey is the real control. Generalize: any modal capture where one hand is committed to the primary input gets the other modality for mode-switching — gesture→key, and conversely a keyboard-heavy capture gets a foot/mouse toggle. The test: if triggering the control would move the thing you're capturing, it's the wrong modality.
Coherent control ranges — bounds must propagate
When one control sets a bound on another (a min, a max, a threshold, an allowed set), the bounded control's UI must reflect the new bound the instant you change it. A "scrub" slider whose min/max attributes drift out of sync with its declared bounds is the most common silent bug — the user moves the bounding slider, nothing visible changes downstream, they assume both are broken.
Rules:
Single source of truth. Hold the bound in state once. Every input that displays it (its own slider, the dependent control's min/max, anything else) reads from that state on every update.
Re-render min/max on every state change. Don't rely on browser-cached attribute values; rewrite them via JS each render. dependent.min = state.lo; dependent.max = state.hi.
Clamp the dependent value into the new range immediately. If the user shrinks the upper bound below the current dependent value, the dependent must snap into range, NOT silently stay outside while the slider shows it pinned to the rail.
No-op regions are slider bugs. If dragging a slider past some value has zero downstream effect (because some other control's bound caps it), that's a coherence bug — either narrow this slider's range to where it actually does something, OR change behavior so it does. Sliders with dead zones train the user to think the studio is broken.
Test by visualisation, not numeric snapshots. Take a screenshot, change the bounding slider, take another. The two must look meaningfully different — or the slider is decorative. A numeric snapshot() showing state changed doesn't prove the pixels did.
Pattern: every time applyState() runs (or its split equivalents), call a syncBounds() helper that walks the dependent-input registry and pushes the live bounds into every min/max/disabled attribute. Clamp values into the new bounds in the same pass.
Paired controls must not cross
A common shape is two sliders that together define an interval — min ⟷ max, near ⟷ far, tightEnd ⟷ wideEnd, start ⟷ end. If the user can drag one past the other, the interval inverts or collapses. Downstream math typically does (x - lo) / (hi - lo) which divides by zero or returns negative t — producing NaN coordinates, collapsed views, or inverted lerps. The user sees the studio "break" but no error fires.
Both ends of the defense:
UI invariant. Keep the two sliders from crossing. On every syncBounds() pass: lower.max = upper.value - MIN_SPAN and upper.min = lower.value + MIN_SPAN (small epsilon, e.g. 2 units, so they can't even touch). The user can't physically drag past the other anchor.
Math invariant. The consuming code (lerp, normalisation, ratio) must guard denominator > 0 and pick a sane fallback for the degenerate case (e.g. clamp t = 1 or t = 0). UI can race the math — always assume the math could be hit with crossed bounds anyway (URL hash, JSON paste-back, programmatic state mutation).
Test the boundary explicitly. When the lookdev exposes both ends of an interval, write a quick check: drag tightEnd to the same value as wideEnd, verify the scene doesn't break. Drag tightEnd past wideEnd, verify same. If you can crash the studio with two slider drags, that's a release blocker.
Architecture
Single-page HTML — <canvas> and/or DOM, vanilla JS, a sidebar of controls. No build step, no framework, no deps unless one is genuinely required. Lives in a project-local scratch dir (e.g. scripts/.lookdev-<name>/ or scripts/.preview-<name>/), gitignored.
Live re-render on every input event. Debounce heavy work via requestAnimationFrame. Keep the loop tight enough to feel like a real slider, not a survey.
Every numeric control is dual-input (MANDATORY): a range slider AND an editable <input type=number>, two-way synced. The drag is for exploring; the typed number is for hitting an exact value (and reading the current one). A static <span> readout is not enough — the user must be able to click it and type. Sync rule: on slider input, write the number field; on number input/change, update state and re-render — but do not overwrite a field while it has focus (guard with document.activeElement), or typing gets clobbered mid-keystroke. Clamp to [min,max] on commit (change), not on every keystroke, so intermediate values like "1" before "12" aren't snapped.
Always include a Reset control that restores every control to its defaults in one click (keep a DEFAULTS object; Object.assign(state, DEFAULTS) then re-render). Cheap to add, and essential once the user has wandered far from baseline.
Always build undo/redo history (MANDATORY). Dialing-in is iterative and lossy — the user will overshoot a good look and need to step back. Bind Ctrl/Cmd-Z (undo) and Ctrl/Cmd-Shift-Z / Ctrl-Y (redo), and surface visible ↶ Undo / ↷ Redo buttons. Snapshot the full serialized state — every control plus any drawn/spatial state (polygons, crop rects, dragged handles, palettes), i.e. the same blob as the settings round-trip (#3), not just slider scalars. Debounce so a continuous drag collapses into one history step (snapshot ~350 ms after the last input, not per event), keep a bounded stack (~100–120 entries), and on a new edit after undo, truncate the redo branch. Restore by re-applying a snapshot through the same applyState path the loader uses (so it can't drift). Guard the key handler when focus is in an <input>/<textarea> so native text-undo still works. A lookdev without undo punishes exploration — the whole point of the tool.
Structured settings round-trip (MANDATORY). Every lookdev MUST expose its full current state as machine-readable, copy-pasteable text — a settings JSON (or equivalent) covering every control, with a one-click Copy button and a visible live readout. This is non-negotiable: the agent cannot bake by eyeballing a screenshot, and the user shouldn't have to describe what they dialed in. The round-trip is: user drags → studio serializes the exact state → user pastes the blob back (or it persists to URL/localStorage) → agent bakes from those literal values with identical math. No control may be tweakable without appearing in the export blob. Mirror the state into the URL query so a look is shareable by link, too.
Reproducible export. Beyond the settings blob, pick by what gets committed:
Copy settings JSON — user pastes back, agent bakes with identical math (port the renderer to Python / build script / etc. and verify the bake matches).
Download asset — page renders the final artifact at full resolution and triggers a download (PNG / SVG / WebP / JSON).
Make exported artifacts re-loadable — sidecar + embedded metadata. When the download is a non-JSON artifact (STL, PNG, GLB, SVG, WebP, video…), the look that produced it shouldn't be strandable. Do BOTH, where the format allows:
Sidecar: download a zip containing the artifact and its settings.json, so the exact state ships next to the result.
Embed the settings inside the file itself, so the bare artifact alone restores the look — then add a drag-drop / file-input loader that reads it back through the same applyState path as the JSON paste. Per-format hooks: binary STL → append MAGIC + uint32 len + JSON after the triangle data (CAM ignores trailing bytes; parse count at byte 80, footer at 84 + count*50) and drop a human note in the 80-byte header; PNG → a tEXt/iTXt chunk; SVG/XML → a <metadata> element or comment; JPEG/MP4 → EXIF/XMP UserComment; GLB → an extras field. The payoff: the user drops last week's STL back on the viewport and the studio re-dials itself — no "which settings made this?" archaeology. Verify the round-trip (export → reset → load → assert state matches) and confirm the artifact still opens in its native tool (the trailing/edge metadata must not corrupt it). Skip only when the format has nowhere safe to stash bytes; the sidecar zip always works as the fallback.
WYSIWYG. The preview frame must match the production context — same background color, same fonts loaded, same container max-width, same object-fit. A generic centered canvas is not WYSIWYG.
Framework-route variant. When the lookdev is for UI layout inside an existing app, build it as a temporary route in the app (app/dev/... or equivalent) so the real components, styles, and tokens are in the comparison. Delete the route once baked.
3D lookdev — orientation gizmo (MANDATORY when the camera orbits)
Any lookdev with a non-fixed camera (OrbitControls, trackball, free fly — anything where the user can spin/tumble the view) MUST include a CAD-style ViewCube in a corner. Free orbit alone disorients: the user loses which way is up, can't get a repeatable canonical view, and can't tell whether they're looking at the front or the back. The cube fixes both problems — it's an orientation indicator and a controller in one. Copy the Autodesk/Fusion 360 ViewCube — that's the interaction users expect; don't invent a different gizmo.
Required behaviour (this is cheap — ~70 lines of Three.js, no excuse to skip):
Live orientation readout. A small second scene/renderer in a corner draws a labeled cube (FRONT/BACK/LEFT/RIGHT/TOP/BOTTOM). Each frame, drive the gizmo camera from the main camera's view direction (gizmoCam.position = (mainCam.position − target).normalize() * d; gizmoCam.up = mainCam.up; gizmoCam.lookAt(0,0,0)) so the cube always mirrors the scene's current orientation.
Click a face / edge / corner to snap (the defining Fusion behavior; it's 26 preset views — 6 faces, 12 edges, 8 corners). Raycast the gizmo and snap each component of the local hit point (|c|>0.55 ? sign(c) : 0) to derive a view direction. One pickable cube then yields faces → ortho views, edges → 45° edge views, corners → iso views from a single mesh — no separate hit zones needed. Animate the main camera to target + dir*currentDist with a short lerp (~0.28/frame), not an instant cut — the motion is what keeps the user oriented.
Drag the cube to orbit freely (also Fusion, also mandatory — the user WILL try to grab it). Use pointer events with click-vs-drag discrimination: on pointerdown record the start and setPointerCapture; on pointermove, once travel exceeds ~4px flip into drag mode and orbit the main camera by the pointer delta (convert the camera offset to spherical around the target, theta -= dx*k; phi = clamp(phi - dy*k, ε, π−ε)); on pointerup, if it never became a drag, treat it as a snap-click. Capture means the drag keeps working when the pointer leaves the little canvas. Cancel any in-flight snap tween when a drag starts.
Roll arrows = Fusion's "rotate". Two curved-arrow buttons (⟲ ⟳) beside the cube that roll the current view 90° about the view axis (rotate camera.up by ±90° around the normalized position−target axis). This is the rotate users mean when they say the gizmo "can't rotate" — drag-orbit is not a substitute for it. Snap-cleanup the rolled up so near-cardinal components land exactly on 0/±1 (keep genuine diagonals). Let it roll in ANY view, including iso — do NOT gate it to face-on views or auto-snap-to-face first. (I tried that "Fusion only rolls in standard views" guard and it backfired: it stops the user rolling an isometric view into the exact orientation they want, which is a primary reason they reach for the arrows. Rolling an iso view is a valid, common move.)
Perspective ⇄ orthographic toggle. Any 3D lookdev should expose a projection toggle. Perspective for a natural read; orthographic for CAD/measure/section work (parallel edges, true elevation, no foreshortening — essential when judging a thickness or aligning a face). Swap by building the other camera, copying position/up/target, and rebuilding controls; size the ortho frustum from the current target distance (h = 2·dist·tan(fov/2)) so the switch doesn't jump scale. Handle resize for both (isPerspectiveCamera → set aspect; ortho → recompute left/right from aspect keeping height).
camera.up + OrbitControls is a TRAP — read this. Three's OrbitControls (r160) captures its orbit-axis quaternion from camera.uponce, at construction. If you mutate camera.up afterward (e.g. to "fix" a top view, or to roll) and leave it, the main-viewport drag silently breaks — OrbitControls keeps orbiting around the old up while the camera renders with the new one. Two consequences for the gizmo: (a) Do NOT flip camera.up for top/bottom snaps. Leave it (0,1,0) and instead nudge the snap direction a hair off the pole (dir = (0,±1,0.0009)) so lookAt with up=+Y doesn't gimbal-lock. (b) When you DO need a new up (the roll arrows), dispose and recreate OrbitControls after setting camera.up, copying target across, so it re-captures the axis. Snaps and Home should reset to up=(0,1,0) and rebuild if currently rolled.
Home / reset-view button beside the cube (Fusion's house icon) that re-frames the object, resets camera.up=(0,1,0), and rebuilds controls if rolled.
Hover highlight the exact zone, not just the face. Fusion subdivides each face into a 3×3 grid — center cell = face, edge cells = edges, corner cells = corners — and lights the hovered cell wrapping across the adjacent faces. Implement with a small pool of up to 3 translucent quads: from the hovered direction d (1/2/3 nonzero axes), for each nonzero axis place one quad on that face at the cell offset (other-axis sign)*⅔. A corner lights 3 quads (one per adjacent face), an edge 2, a face 1. A plain whole-face tint is wrong — the user can't tell a corner-pick from a face-pick. Also set a grab/grabbing cursor so the cube reads as draggable.
Orient the model so FRONT is the face the user cares about. The cube's labels are fixed to world axes, so how you place the model decides what "FRONT" shows. For a relief/panel/anything with a hero face, stand it so the hero face points world +Z (= FRONT) and image-up points +Y — don't lay it flat facing +Y, or FRONT shows a meaningless edge and TOP shows the hero (surprising and "wrong" to the user). Watch the displaced-axis sign too: Three's PlaneGeometry pushes -y, so vertex row 0 is +Y (top) — map image row 0 (top) to it with no flip, or your relief comes out upside-down. Verify by snapping FRONT and eyeballing against the source image; don't trust the index math.
Build solids, not floating sheets. A displaced PlaneGeometry is a single hollow surface — fine for a quick look, wrong the moment the user inspects it. In X-ray (or any side view) the raised bumps read as hollow domes floating above the base with a gap, and it's not watertight for STL/CAM. If the thing is a real object (relief, terrain block, carved panel), build a solid heightfield: displaced top surface + perimeter skirt walls + flat bottom, so it's rooted on its base. The user will notice "the back doesn't touch the backplate." Set the material DoubleSide so hand-wound walls never render black.
Section / X-ray for hidden internal dimensions. When a control sets something you can't see from outside — wall thickness, a backing/backplate, internal clearance, draft — add an X-ray/section toggle so the user can actually see what they're dialing. Cheapest version: ghost the outer shell (transparent, opacity~0.15, depthWrite:false) and render the measured solid (the backplate slab, the remaining wall) as an opaque distinctly-colored mesh with a bright edge line at the critical boundary; pair it with a side ortho snap so the dimension reads as a clean band. (A true clipping-plane section with caps is the fancier version; usually not worth the stencil work.) Don't make the user infer a hidden thickness from a number alone when one toggle can show it.
Keep it in world/view space aligned to how the model is displayed (account for any root rotation you applied). Verify by visualization, not math (these all bit me): click TOP then drag the main viewport and screenshot — confirm it still orbits (catches the camera.up trap); hover a corner and screenshot the cube — confirm the corner zone lights across faces, not the whole face; click a roll arrow from an iso view and screenshot — confirm it snaps to a face (not a diagonal roll); snap FRONT and confirm the hero face is upright. Genuinely-optional Fusion extras: the adjacent-face triangle arrows (drag-orbit covers them), the N/E/S/W compass ring, and the right-click "set current view as Home" menu — skip unless asked.
Workflow
Build the studio for the specific question. Don't make it generic. If the user is choosing a hero crop, the studio shows the actual hero. If they're choosing a font, the studio is reading sample text.
Serve locally. Never hardcode a port — bind a static server to port 0 (the OS hands back a free port) for static HTML, or use the project's dev server for framework routes. Give the user the URL.
Verify it works headlessly before handing it over (headless Playwright). Don't ask the user to debug your scaffolding.
User iterates. They paste back a settings JSON, click a Download button, or say "go with N" / "use this".
Bake. Render the chosen state into committed assets / production code with reproducible math. Verify the baked result matches what they dialed in (a quick screenshot diff is fair).
Tear down the scaffolding. Delete the lookdev dir / dev route — it was decision-time scaffolding, not production code. Commit + deploy.
Anti-patterns
Static N×M comparison grid — limits the user to your guesses; takes longer than a switcher; doesn't give them the in-between point they actually wanted.
Numeric prompt before the slider — "what saturation do you want?" is the wrong question; let them drag.
Numeric inputs for spatial decisions — drag the element. Sliders for opacity, drag handles for position.
Drift between preview math and bake math — when both JS preview and Python bake exist, port one to match the other and verify on a known input.
Building inside production routes — keep scaffolding isolated and trivially deletable. Reach for app/dev/... then nuke it.
Skipping the WYSIWYG details — preview without the real font / container / background lies to the user.
No structured way to read the state back out — a studio with no copy-able settings blob forces the agent to bake from a screenshot and the user to narrate values by hand. Every control must round-trip through a machine-readable export (see Architecture #3).
Handing back a wall of prose for "review" — pasting a long doc/blog into chat (or shipping the markdown file) and asking the user to react is NOT lookdev. For any document/copy/media review, build the text & media annotation studio (direct edit + highlight + comment + media-flag) so the user marks up the rendered artifact and the markup round-trips back as a patch. A boring text dump the user has to read and reply to in chat is the exact thing this skill exists to replace.
Working example
A worked example — an image-treatment studio:
extracts a Lab-k-means dominant palette with coverage %, exposes sliders
(resolution, colorize, saturation, gap, glyph, contrast), per-band color
pickers, a luminance-vs-nearest mapping toggle, a Copy-settings-JSON
button, and a --bake-json Python path that renders the chosen state to
committed PNG/WebP with math identical to the JS preview. The preview
canvases match the production thumb and hero shapes exactly.
Related (the studio / narrative family)
lookdev is one of two flagship narratives — human-in-the-loop (you, the human, judge and
tune). Its determinism-narrative sibling is deterministic-design (render → measure the
UI, numbers not vibes). The family it chains with:
deterministic-design — the other flagship; measure/judge design output deterministically.
screenstudio-alternative — human-in-the-loop video/demo polish studio (NLE timeline).
macos-screen-recorder — capture a studio session or demo (display + system audio).
lookdev-auto — the automated counterpart: a vision model judges instead of you.
The foil to lookdev's thesis — use when there's no human to sit the loop.
Limitations
Lookdev is useful only when the user can inspect or mark up rendered variants; it is overkill for small deterministic edits.
A studio must faithfully mirror production fonts, media, containers, and constraints, otherwise the chosen settings can be misleading.
Human preference remains the source of truth, so the workflow cannot guarantee a universally "best" design or media treatment.
