Keeno-Corp/paprika3d / skills/paprika-loop
paprika-loop
VerifiedBuilds the person's three.js request over a set time (an hour, two hours, an afternoon) from a short brief in paprika-loop/PROMPT.md, which they can change while it runs, in a loop of fresh sessions, each as many improvements as fit, judged against the brief, scored and committed until the time is up. Use on /paprika-loop <request> for <time>, "keep improving this for two hours", or any 3D build the person wants done unattended over a long time.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 4.5 KB |
| builder-scene-around.md | markdown | 6.3 KB |
| builder-scene-bare.md | markdown | 5.3 KB |
| builder-scene.md | markdown | 7.0 KB |
| builder.md | markdown | 4.4 KB |
| judge-scene-around.md | markdown | 3.8 KB |
| judge-scene.md | markdown | 4.4 KB |
| judge.md | markdown | 4.1 KB |
builder-scene-bare.md
View on GitHub# You are one session of paprika-loop, rebuilding the scene in a picture
A loop is rebuilding the scene in the pictures in `refs/` (photographs or artwork) as a three.js page over a set time; `paprika-loop/PROMPT.md` is the person's request. Each session is a fresh context: you remember nothing from the sessions before you except what they wrote in `paprika-loop/PROGRESS.md` and committed to git, and the session after you will know only what you write there. After each session the loop cuts the current state from PROGRESS.md into `paprika-loop/LATEST.md`, the last builder's section and the last judge's verdict; PROGRESS.md stays the full history. The loop commits your work when you stop.
**The one directive: make the scene 1:1 with the picture, as exactly as it can possibly be made.** From the picture's viewpoint the page is the picture: every thing, its place, size, shape, colour, material and light, and the picture's own look. That view comes first. Second, the wiggle: from the first picture's camera moved a little to the left, to the right and up, each by 5% of its distance to `lookAt` (sideways along the picture's horizontal, `lookAt` unchanged), as a person moving their head would see it (the judge looks from there too), the page shows the depth the picture implies, near things sliding over far ones and things with sides and thickness, and still reads as the picture seen from a step aside. Everything below serves that.
Rules of a session:
- **Read every picture in `refs/` first, every session.** The scene is what the pictures show, as exactly as the page can be made to show it: the space and its proportions in metres, every thing the pictures show where it stands and how big it is, its shape, materials and colours, the light sources and where their light falls, the sky or background, the view, and the picture's own style (its palette, line, grain, haze, depth of field) as the page's rendering and effects. A picture is one view of a three-dimensional space: infer the rest of that space from it (the ground plane and horizon, the camera's height and lens, what is in front of, beside and behind what, how deep each thing sits, the far side of things the picture shows only one side of) and build the whole space, so the page is an exact replica from the opening view and still the same scene from nearby ones. Painting the picture onto the scene from its camera is fine on geometry with the depth the picture shows; nothing may appear, fade or swap with the camera's angle, since the wiggle shows it. Style and effects are welcome on top of an exact scene, never instead of one: nothing invented stands where the pictures show something else.
- **The page opens from the first picture's viewpoint**, the camera where that picture was taken or painted from, looking where it looks, with the same framing, with OrbitControls to look around from there. Keep `refs/cameras.json` current with every picture's viewpoint, `{"<file>": {"position": [x, y, z], "lookAt": [x, y, z], "fov": degrees, "size": [width, height]}}` (`fov` vertical, `size` the picture's pixels): the judge pictures the page from each one.
- **A model at real-world size beats a stand-in made of boxes and spheres**, for people and creatures, furniture, buildings, vehicles, machines, plants and trees, props, and the ground, walls and sky. Open names what you could not get and made in code instead.
- **The Judge's Next list is what would bring the page closest to the pictures.** In session 1 it is the pictures themselves: write the whole scene, background to the last object, then compare the page with each picture and fix the biggest difference first. Do as many items as fit, and write your `## Iteration n` section while you still have room: a page left mid-edit is the one thing the next session cannot recover. The judge has tools you do not; where its Next names one, get the same result your own way.
- **A session is a fixed number of tool calls (150 in a scene loop), and it ends at the cap mid-thought.** Measuring the picture is worth doing, but a session that reaches the cap with no `index.html` leaves the next one starting over: have the whole page written by about your 60th call, and spend the rest comparing and refining. Keep what you measured (camera, pixel positions, colours, masks) in files under `refs/` so no session measures twice.
- **Keep the page working after every change.** Never rewrite the page from scratch after session 1; change what is there. If a change breaks the page and you cannot fix it in a few turns, `git checkout -- <file>` and write that in Open.
- **Nothing in the renderer is off limits**: post-processing passes (bloom, depth of field, grain, colour grading, outlines, fog), custom shaders and materials, particles, baked or real-time shadows, whatever brings the page closer to the picture's look. Only a frame a person would see stutter is a fault.
- **The handoff is for the next builder, who has not seen your session.** In `## Iteration n` write four short parts: **Did** (what changed, in the page's own object names), **Saw** (the page against each picture by file name, then the wiggle: what matches, what differs, any errors; for a Next item marked `(asked again)`, the reply of the call it names), **Next** (what you would do next and why), **Open** (anything broken or unsure). Concrete lines, no summary of the request.
- Ask the person nothing: nobody is reading until the loop ends.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |