Josh Creator Scene Skill

Build a primary scene that is credible in the user's actual room, not a generic pretty render. The product promise is: upload one real room image, see the recommended talking-head result in that space, then receive exact setup instructions. Keep two distinct alternatives available when comparison would materially help, but do not make the number of versions the headline. The user should not need to understand prompts, CLI, SDKs, lighting theory or set-design terminology.

Core rules

• Treat the user's photo/video as the source of truth for room geometry, fixed openings, desk orientation, and usable depth. State assumptions instead of inventing architecture.
• Never say only "left" or "right". Use 画面右（人物左）, 画面左（人物右）, or 相机右 as appropriate. Define the coordinate system once in every practical plan.
• Use only genuine A-roll examples: the creator is visibly speaking to camera in the featured setup. Do not substitute B-roll, course slides, product photos, studio tours, stock interiors, or unrelated screenshots.
• Present generated visual effects as a concept preview, not as proof that the proposed objects fit or that lighting values are exact.
• Keep the creator visually dominant. Background is for depth, identity, and proof of work—not clutter.
• Preserve a supplied person’s identity, pose, room planes, camera angle, and architectural constraints when generating a mockup. Do not beautify the face or alter body, room dimensions, windows, doors, or desk direction.
• Maintain an approval ledger throughout the conversation. Record elements the user explicitly liked, rejected, or corrected. Approved elements outrank a newly selected inspiration image; rejected elements must not reappear.
• Distinguish existing architecture, plausible movable additions, and imagined construction. The user may approve a flush shelf, plant, practical lamp, art, or microphone even when these are not in the source photo. Do not interpret "preserve the room" as "leave the room empty."

0. Intake and spatial read

Accept one scene photo as the minimum viable input. If available, also collect a sweep video, intended platform and aspect ratio, content lane, budget, existing gear, rental/installation constraints, and anything that must stay.

Do not block on missing details. Make a best-effort spatial read, mark uncertain assumptions, and request only the one extra input that materially changes the solution—for example, a 10-second room sweep when depth behind the chair cannot be seen.

Before researching, summarize:

1. camera position and likely framing;
2. subject position, eyeline, desk edge, and any real depth path;
3. foreground/midground/background planes, fixed visual noise, practical power locations, and daylight control;
4. the scene's suitable brand direction in one sentence.

Create the approval ledger from the whole task context:

• Keep/approved: room geometry and any prior visual elements the user explicitly endorsed.
• Avoid/rejected: screens, furniture, layouts, colors, or lighting treatments the user disliked.
• Unknown: choices that may vary across the three versions.

Read spatial-reading.md for the checklist and coordinate convention.

1. Find three reference creators

Use agent-reach for online discovery and follow web-access requirements for any live web or browser operation. Announce the research capability before using it. Before research, run agent-reach doctor --json; if the command is not on PATH, also try $HOME/.agent-reach-venv/bin/agent-reach doctor --json. Prefer direct creator/video/profile sources over reposts.

If Agent Reach is genuinely unavailable, use the environment's official web/search/browser capability as a declared fallback. Never fabricate creator names, URLs, or A-roll frames. If no live network path exists, finish only the spatial read and explain the missing dependency instead of pretending Stage 1 is complete.

Read reference-research.md. Find exactly three candidates before making the room-specific designs. Every candidate must include:

• creator name, primary platform/profile or source-video URL, lane, and a captured/linked A-roll frame;
• the exact visual elements that match the user's room;
• a match score with reasons, plus what must not be copied;
• a one-line lighting read: key direction/softness, background practical, separation, and contrast.

Choose three different creators. Optimize for physical applicability rather than follower count or aesthetic novelty. If no source has a reliable public A-roll frame, say so and replace it with a better-evidenced candidate.

Use the references as design evidence. Unless the user explicitly asks to stop at raw references, do not make them choose yet; users cannot reliably judge a reference until they see it translated into their own room.

2. Convert references into a primary direction and alternatives

Read set-design.md. Create one recommended primary direction plus two distinct, credible alternatives. Each may map to one reference or combine a reference's lighting grammar with approved elements from a prior user-liked render. Translate each direction through these layers:

1. Composition: final crop, camera height, subject offset, visible depth path, and negative space.
2. Set: retain/use/remove/add; use only objects that fit real planes and do not create a fake floating desk or display.
3. Lighting: key, negative fill, practical/background light, optional rim; specify direction in both coordinate systems.
4. Brand: color/material hierarchy for the creator lane. For an AI/tech creator, favor controlled contrast, restrained warm practicals, dark neutral clothing options, purposeful tools, and a quiet proof-of-work background.
5. Approval ledger: retain approved elements and exclude rejected ones in all three directions.

The three versions must differ in a meaningful visual decision, not merely color temperature. Recommended territories:

• A — Warm creator studio: strongest lifestyle depth and practical-light warmth.
• B — Minimal product designer: fewer objects and cleaner negative space, while retaining a real midground.
• C — Dark professional tech: deeper contrast and stronger separation, without RGB/gaming cues.

Do not label one as the cheap or inferior version. All three should be desirable and physically achievable.

3. Show the visual effect

Use the imagegen skill and its built-in image tool to edit the supplied scene image. Read mockup-prompting.md before prompting.

Generate the recommended primary mockup first, then generate two alternatives using separate image-generation calls when the user asked to compare or when the direction is genuinely uncertain. Preserve the source room’s actual geometry and the person's identity in every version. Use references for mood and visual grammar, not literal copying. If the user supplies a previously approved render, include it as an approved visual reference with higher priority than the creator inspiration.

Inspect every generated result against the source and approval ledger. Regenerate a version once if it changes room structure/camera orientation, loses the approved midground, reintroduces a rejected element, or adds objects that cannot fit. Present the primary result prominently. Keep alternatives secondary and explain only the meaningful visual decision that differs.

Ask the user to confirm the primary direction. If they reject it or request comparison, show the alternatives and let them choose.

4. Deliver the executable setup

Read execution-deliverable.md. Give an implementation-ready plan containing:

• top-down placement with camera, subject, desk, all light positions, and the real background/depth path;
• camera recipe: lens/field of view, height, distance, crop, eyeline, and exposure/WB starting point;
• lighting recipe: modifier, relative position, height, angle, distance, color temperature, starting output, what it illuminates, and the adjustment test;
• staging list: exact objects, positions, scale, color/material role, plus items to remove/hide;
• budget tiers and a purchase list limited to missing necessities;
• a 10-minute setup sequence and a phone-preview correction loop.

Use relative brightness first (face, background, practical) rather than false precision. Add absolute power/lux only when the fixture or meter is known. Explain every visual move in terms of the user's content identity and the chosen reference.

Finish condition

The task is complete only when the user has: verified reference evidence, one confirmed room-specific visual direction, and specific instructions they can set up alone. Offer alternatives when useful, and offer a final calibration pass from one test frame after they build it.