game-assets

Files (1)

• SKILL.md

SKILL.md (first lines)

# Game Assets — Planning, Generation & Processing

## Overview

Three-phase asset workflow:

1. **Plan** — analyze game, identify assets, budget allocation
2. **Generate** — create images, 3D models, animated sprites
3. **Process** — background removal, frame extraction, loop trimming

---

## Phase 1: Asset Planning

### Input

- `budget_cents` — total budget (or remaining for iterations)
- Game plan spec (in `.specs/`) — game description, risk tasks, main build with asset needs
- Architecture notes (draft in `.ai.tmp/`) — architecture with Asset Hints section
- `reference.png` — visual composition target
- Art direction notes (draft in `.ai.tmp/`) — art direction from visual target phase

### Workflow

1. **Read `reference.png`** — understand visual composition, proportions, environment layers.
2. **Read architecture Asset Hints + game plan Assets needed** — reconcile both lists (they may overlap or complement each other).
3. **Categorize assets:**
   - **3D models** — characters, vehicles, key props, buildings (need geometry)
   - **Textures** — ground, walls, UI backgrounds (flat, tileable)
   - **Backgrounds** — sky panoramas, parallax layers, title screens (large scenic)
   - **Animated sprites** — characters/objects with multiple actions (plan motion graph first)
4. **Budget allocation** (see cost table below).
5. **Identify anchors vs derivatives** — anchors generated first, derivatives use anchor as `--image` input.
6. **Generate, review, convert** — images in parallel, review PNGs, retry bad ones (max 1), convert approved 3D refs to GLBs.

### Cost Table

| Asset Type                 | Backend             | Cost               | Notes                           |
| -------------------------- | ------------------- | ------------------ | ------------------------------- |
| Texture / simple sprite    | Grok                | 2c                 | Fast, simple images             |
| Character / reference      | Gemini 1K           | 7c                 | Precise prompt following        |
| Background (simple)        | Grok                | 2c                 | Sky, clouds, abstract           |
| Background (precise)       | Gemini 2K           | 10c                | Specific layout/composition     |
| 3D model (full)            | Gemini 1K + GLB     | 47c                | 7c image + 40c GLB high quality |
| Animation reference        | Gemini 1K           | 7c                 | Once per character              |
| Animation action (root)    | Gemini pose + video | 7c + 5c x duration | From reference                  |
| Animation action (chained) | Video only          | 5c x duration      | From predecessor's last frame   |

Reserve ~10% of budget for retries. Prioritize by visual impact — cut low-impact assets first.

### Backend Selection

- **Gemini** (`--model gemini`) — reference images, character design, 3D model references, animated sprite refs/poses, backgrounds with precise layout. Costs more but reliably matches prompt.
- **Grok** (`--model grok`, default) — textures, simple objects, item kits, props, simple scenic backgrounds. High-quality but imprecise — great when exact prompt adherence doesn't matter.

### Art Direction Usage

Read the **Art direction** from your draft but do NOT mechanically prepend it. Different asset types need different treatment:

- **Textures** — often need no style language at all
- **3D model references** — need clean studio lighting; style cues can hurt mesh quality
- **Backgrounds** — benefit most from art direction language
- **Sprites** — some style cues, adapted to the subject

### Image References for Consistency

Feed a generated image as `--image` input when subsequent assets need to match it:

- **Style family** — one hero asset as input for the rest of the set
- **Multiple views** — front view as input for side, back, 3/4 angle
- **Variants** — base object as input for recolors, damaged versions, sizes
- **Scene coherence** — background as input for foreground props

Generate anchors first, review, then fan out derivatives in parallel.

---

## Phase 2: Generation CLI

All tools are TypeScript-based. Run from project root.

### Generate Image

```bash
godot-asset-gen image --prompt "the full prompt" -o assets/img/car.png
```

Options:

- `--model` (default `grok`): `grok` (2c), `gemini` (5-15c by size)
- `--size` (default `1K`): Grok: `1K`, `2K`. Gemini: `512`, `1K`, `2K`, `4K`
- `--aspect-ratio` (default `1:1`): `1:1`, `16:9`, `9:16`, `4:3`, `3:4`, `3:2`, `2:3`
- `--image` — reference image for image-to-image editing

Typical combos:

- `--model gemini --size 1K` — refs, characters, 3D refs (7c)
- `--model gemini --size 2K --aspect-ratio 16:9` — backgrounds (10c)
- `--model grok` — textures, simple objects (2c)

### Generate GLB (3D Model)

```bash
godot-asset-gen glb --image assets/img/car.png -o assets/glb/car.glb
```

Options:

- `--quality`: `default` (50c) or `high` (40c, HD textures)

Do NOT remove background before GLB conversion — Tripo3D needs the solid white bg.

### Generate Video (Animated Sprites)

```bash
godot-asset-gen video --prompt "walking right, smooth cycle" --image assets/img/pose.png --duration 2 -o assets/video/walk.mp4
```

Options:

From dragoscirjan/opencode-config, MIT.

More from dragoscirjan/opencode-config