Load this from new-work.md whenever any image generation is available, a harness-native tool or the API fallback context.mjs reports. PRODUCT.md and DESIGN.md are preconditions. New-work has already resolved the visual world; this file must not reopen it.
The purpose of a probe is to test composition, narrative, hierarchy, density, focal moment, signature use, and image requirements. It is not a second identity workshop. Keep DESIGN.md's palette, typography direction, material language, component character, imagery stance, and motion grammar fixed.
Render three distinct high-fidelity north-star comps of the requested surface, with whatever generation capability exists, saved under .impeccable/mocks/ so they survive the session. Comp at the surface's own viewport: portrait at device size for a native app or mobile-first surface, desktop landscape otherwise; a phone screen comped landscape misstates the composition before anything gets built against it. Comps are the build thread's own work, never delegated: the thread that writes the comp prompts holds the direction's full context, and it has already seen every comp when the build starts. Open every image you produce or reference by its workspace-relative path, never an absolute one: sandboxed viewers reject absolute paths, and everything under the project root has a relative path. Base them on the real content and the surface concepts already developed with the user. Three is the number: one comp invites rubber-stamping, and the spread between three is what surfaces the composition worth building. A decision-page sketch is not a probe: it chose the direction at deliberately unfinished fidelity, so the three comps render regardless, and the chosen card's sketch seeds at most one of them.
Treat each comp as a direction test, not a screenshot specification. Core UI text, responsive behavior, accessibility, semantics, and interaction states remain implementation responsibilities.
Show the three together: in the harness when it can display images, otherwise on the decision page (serve-question.mjs, one option per comp with the comp as its hero). Ask what should carry forward, what feels false to the world, and whether the selected surface concept should be approved, combined, revised, or rejected. Then stop and wait. A structured simulated user counts as attended and receives the same question.
Do not begin code until the user approves a direction or explicitly delegates the choice. If they delegate, choose using the task brief, PRODUCT.md, and DESIGN.md, and state the evidence. Approval refines the task concept; it does not modify DESIGN.md.
This approval point has no substitute and no skip condition. When the structured question tool errors, fall back to the decision page; only after both fail may you treat the choice as delegated, and a delegated pick is still recorded exactly as an approval is and disclosed in your first reply, not your last. The finish reviewer treats a build with generated comps and no recorded approval as carrying a material finding.
After approval, record the choice where tools can find it: the approved comp's path goes in the surface brief, and the approved comp's .json prompt sidecar gains "approved": true (every comp generated through generate-image.mjs has one; create it if a native tool didn't). The sidecar travels with the mocks folder, so the approval survives sessions and machines that never see the brief. Then summarize the composition and the parts of the comp that must not be literalized, return to new-work.md, record the direction contract from the approved surface concept, and build.
Before building, read the approved comp as a design system and record it in the brief: component grammar, corner language, line weights, elevation treatment, and the type ramp, because everything the comp does not show gets built from this record, and without it the fallback is the model's stock kit of square boxes, 1px grids, bento cells, and hard shadows. Then inventory the comp's major visible ingredients in writing (a short table in the surface brief or working notes; the finish reviewer audits shipped assets against it) and choose an implementation medium for each: semantic HTML/CSS/SVG, existing project asset, generated raster, sourced raster, icon library, canvas/WebGL, or accepted omission. The same written inventory names the comp's compositional commitments: navigation items and icons, headline levels and their scale relationship, signature geometry such as seams, masks, and overlaps, and each section's arrangement and density. The primary action gets its own row with its own medium: when the comp dissolves, stamps, erodes, or otherwise physically works the main CTA, that treatment is signature material on the page's most important element, and shrinking it to a border trick or a few decorative pixels is the compliance-token version of commitment. An element never written down is the element the build silently drops, and the direction contract's 150 words cannot carry this list, so this inventory is where it lives.
The medium column is where an approved design most often dies, so it obeys a gate: the medium is decided by what the comp region shows, never by what feels buildable in the current stack. A human figure, a product object, machinery, or any material with lighting and depth is raster whatever the stack, and so is any texture by that name alone: woven cloth, paper grain, fabric, leather, brushed metal need no depth argument, because a CSS gradient or layered background is not a texture medium and "layered CSS textures" is not a medium at all. Writing "silhouette" for a photographic figure, or "CSS" for a sculpted panel's finish or a cotton field's weave, is not a medium choice, it is the quiet deletion of the approved design, and it is how a comp full of physical material becomes a flat page with the same section order. Style does not move this boundary: a comp region with perspective, shading, figure drawing, or dense mechanical detail is illustration however line-drawn it looks, and no build session can author illustration as vectors, so it regenerates as raster like any photograph. Authored SVG covers what a session can specify exactly, diagrams with countable elements, controls, flat shape systems, and it ends where drawing skill begins; an instruction-manual world does not convert its illustrations into diagrams, it makes them line-art illustrations. Produce such regions by regenerating them cleanly, with the approved comp and its embedded prompt as the reference for a fresh render at asset resolution; never crop pixels out of the comp itself, whose effective resolution sits far below asset grade. Dropping an image-native region instead of producing it is a scope decision the user makes at the approval point, never a silent flattening after it. Generated imagery is a material, not a claim: evidence rules bind assertions, specs, testimonials, and photographs presented as real, never render fidelity, so "no photography on hand" forbids fake proof, not an illustrated hero.
The gate runs both ways: precise geometry, hard-edged shape systems, diagrams, expressive motion, shaders, and anything interactive are vector and GPU territory (SVG, canvas, WebGL), where a raster flattens what should move, scale, and respond, and code executed safely and professionally remains first-class there. A field or texture built from many small elements carries a quantity commitment either way: write down its approximate density and coverage ("thousands of glyphs over two-thirds of the fold, dense at the top fading into the path"), because a field rebuilt at a tenth of its density passes every checklist and still is not the design. TYPE rows carry the same discipline: name the face's compression class, and render one headline word against the comp before building on it; a visibly wider or lighter silhouette means the face is wrong, and every section built on it inherits the miss. Raster is for what the world paints; code is for what the world draws, animates, or reacts with, and choosing code there is ambition, not economy. Every produce entry is produced before the build ships, through the asset producer or in the current thread; an inventory with unproduced entries is an unfinished build, and this gate is where imagery-free pages come from when it is skipped.
Pay special attention to the dominant composition, signature use, image-native content, second-fold system, and any interaction the still image only implies.
Treat the comp as a north star, not something to trace, and know what that allows: translation into semantic, responsive, accessible code, never recomposition. Keeping the palette and mood while redrawing the topology is a second art direction, not an adaptation. Do not rasterize core UI text or controls. Do not substitute a different visual driver after approval without asking.
Generation context is part of the asset: a build composed by a thread that never saw the prompts places assets it does not understand. So prefer generating build-critical imagery in the build thread when the budget allows, and when a subagent produces assets instead, every asset must carry its prompt, and the builder reads those prompts before composing a single one of them. The carrier is uniform across harnesses: after generating any image with any tool, native or generate-image.mjs (which does it automatically), run node .claude/skills/impeccable/scripts/embed-prompt.mjs <image> --prompt "<the prompt used>" so the intent lives inside the file itself and survives copies between machines and harnesses; --read recovers it from any impeccable-generated image.
When the harness runs subagents, spawn the shipped asset producer every time, even when the inventory's produce bucket looks empty: its manifest is the independent second opinion on your media, and runs that skipped the spawn are the runs whose cotton became CSS. An honestly empty manifest costs one cheap spawn; a wrongly empty produce bucket costs the build its materials. Use the producer, impeccable-asset-producer (impeccable_asset_producer in codex; /impeccable-asset-producer in Cursor; on GitHub Copilot say "Use the impeccable-asset-producer agent"): give it the approved comp, output paths, required dimensions and formats, transparency needs, crop notes, and what must remain semantic code. Otherwise produce the minimum required assets in the current thread by the book: load degraded/asset-producer.md and follow it inline, with whatever generation exists, the native tool or generate-image.mjs.
Convert images with a converter context.mjs reported at boot (the IMAGE_TOOLS line); probe only when it reported none, at most once per session, never per image.
Return to new-work.md for the direction contract, implementation, and the finishing pass.
