{"id":"765edc30-436d-4ea0-8ab4-a8f6534e5309","slug":"fun-skills","kind":"skill","name":"fun-skills","summary":"Generative SVG interior scenes & hand-inked sketchnotes via deterministic physics evals, metric arithmetic & self-improving agent loops","owner":"surendranb","repoUrl":"https://github.com/surendranb/fun-skills","license":"MIT","licenseTier":"permissive","trust":"community","stars":0,"match":null,"fewStars":true,"format":"skill","createdAt":"2026-09-02T16:31:22.000Z","addedAt":"2026-10-21T10:24:54.824Z","logoUrl":"https://avatars.githubusercontent.com/u/3054895?u=0649445cad6d845760c491959b386c559874c00a&v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["design-media"],"tags":["ai-agents","antigravity","evals","generative-art","llm-evals","self-improving-agents","sketchnote","spatial-computing"],"runtime":"python","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-23T04:14:10.595Z","sample":false,"url":"https://conduits.dev/tools/fun-skills","language":"Python","transports":["stdio"],"install":{"pypi":"fun-skills"},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"# Fun Skills: Spatial & Doodle Architects\n\n

\n \"License:\n \"Python\n \"Physics\n \"Zero-JS\n \"Pure\n

\n\n> Generative SVG architectural interiors and hand-inked sketchnotes engineered through **deterministic physics evals**, **metric perspective arithmetic**, **kinematic motor curves**, and **self-improving feedback loops**.\n\n---\n\n## Table of Contents\n\n- [The Core Innovation: Deterministic Evals > Vague Prompting](#the-core-innovation-deterministic-evals--vague-prompting)\n- [The Showcase: One Brief, Two Architects](#the-showcase-one-brief-two-architects)\n - [1. Spatial Architect (Kyoto Kissaten)](#1-spatial-architect-kyoto-kissaten)\n - [2. Doodle Architect (Kyoto Kissaten)](#2-doodle-architect-kyoto-kissaten)\n- [How to Install & Use with Your AI Agent](#how-to-install--use-with-your-ai-agent)\n - [Google Antigravity & Gemini CLI](#google-antigravity--gemini-cli)\n - [Claude Code & Claude Desktop](#claude-code--claude-desktop)\n - [Cursor, Windsurf, or Any LLM Harness](#cursor-windsurf-or-any-llm-harness)\n- [Running the Deterministic Evaluation Suites](#running-the-deterministic-evaluation-suites)\n- [Repository Structure](#repository-structure)\n- [License](#license)\n\n---\n\n## The Core Innovation: Deterministic Evals > Vague Prompting\n\nLLMs struggle with spatial reasoning, perspective foreshortening, and authentic motor kinematics when prompted naively. They generate floating clip-art, impossible lighting angles, and sterile geometric boxes.\n\n**Fun Skills solves this by pairing generative skills with deterministic evaluation engines:**\n1. **Mathematical Grounding**: Every scene begins with declared metric scales (`288px = 1m`), floor datum planes (`y = 720`), and standing camera heights ($H_{\\text{cam}} = 1.65\\text{m}$).\n2. **Deterministic AST Evals**: The evaluation script parses the raw SVG DOM to enforce 31 non-negotiable physical laws (emitter-to-pool optics, span closures, contact shadows, collision prevention).\n3. **Pixel-Tier Luminance Histograms**: Headless Chrome renders the output and verifies that shadows reach deep tones, highlights pop, and zero pure black (`#000000`) is used.\n4. **Self-Improving Loops**: The agent runs the eval, receives structured programmatic error diffs, and corrects the SVG code until it reaches 0 errors.\n\n---\n\n## The Showcase: One Brief, Two Architects\n\nWe asked both skills to render the exact same real-world scene:\n> *\"Kyoto Kissaten: Gion district morning drip coffee counter with Shoji window looking out to bamboo, Hinoki bar slab, copper gooseneck kettle, triple brass V60 drip station, calligraphy menu slats, and shop cat.\"*\n\n---\n\n### 1. Spatial Architect: Kyoto Kissaten\n\n![Kyoto Kissaten Spatial Progression Grid](examples/spatial/01-kyoto-kissaten/kyoto-kissaten-progression-grid.png)\n\n#### 8-Shot Progression & Physics Eval Loop\n\n| Shot | Visual State | What the Eval Pointed Out | How We Fixed It |\n|---|---|---|---|\n| **01** | \"Spatial | **FAIL (3 errors):**
• `datum_declared`: Missing floor datum
• `scale_declared`: No metric scale
• `camera_height`: Ungrounded floating coordinates | Stated Stage 0 metric brief: declared `` and `` reference plane. |\n| **02** | \"Spatial | **FAIL (2 errors):**
• `room_declared`: Missing 3D room boundary
• `envelope_closes`: No foreshortened floor planks | Declared ``. Built Phase 1 envelope: plaster back wall (`y=0..720`), foreshortened hinoki floor (`y=720..900`), and cedar baseboard. |\n| **03** | \"Spatial | **WARN:**
• `wall_collisions`: Fixtures colliding with aperture frame
• Missing garden depth cue | Ordered Phase 2 Apertures before Phase 3 Fixtures. Added shoji window frame with misty bamboo silhouettes and cedar menu slats. |\n| **04** | \"Spatial | **ERROR `pool_shape_mismatch`:**
• *Window aperture projected to a round ellipse pool instead of a quadrilateral.* | Replaced ellipse with Law F3 floor-projected quadrilateral polygon: `` aligned to 34° morning sunbeam. |\n| **05** | \"Spatial | **ERROR `object_scale`:**
• *Counter height eyeballed rather than derived from metric scale.* | Derived exact 0.80m bar height ($230\\text{px}$) from $288\\text{px} = 1\\text{m}$. Added `data-real=\"2.64x0.72\"` to solid hinoki counter casework in Phase 5. |\n| **06** | \"Spatial | **WARN `density`:**
• Flat prop surfaces lacking Gibsonian physical affordances | Added Phase 6 artisanal pour-over gear: copper gooseneck kettle, triple brass V60 station, ceramic cones, amber bean jars, matcha chawan, and cup on saucer. |\n| **07** | \"Spatial | **ERROR `no_pure_black` & `floor_contact`:**
• Pure `#000000` black shadows
• Missing contact occlusion at datum | Replaced black with warm indigo `#2c1d11` and cedar `#5c4033` contact shadows beneath counter slab, along baseboard, and under props at datum `y=720`. |\n| **08** | \"Spatial | **PASS (0 errors, 0 warnings):**
• Pixel-tier histogram verified
• All 31 physical rules satisfied | Grounded bentwood stools on datum `y=720` (`data-real=\"0.24x0.45\"`), added curling steam wisps, and applied zero-JS GPU architectural paper grain shader. |\n\n---\n\n### 2. Doodle Architect: Kyoto Kissaten\n\n![Kyoto Kissaten Doodle Progression Grid](examples/doodle/01-kyoto-kissaten/kyoto-kissaten-doodle-grid.png)\n\n#### 8-Shot Progression & Kinematic Inking Loop\n\n| Shot | Visual State | What the Eval Pointed Out | How We Fixed It |\n|---|---|---|---|\n| **01** | \"Doodle | **FAIL (4 violations):**
• Rigid `` & `` primitives
• Monotone `#000000` black strokes
• Zero motor kinematics or human hand variation | Rejected rigid primitives. Established freehand composition with organic coordinate spacing. |\n| **02** | \"Doodle | **WARN:**
• Unbalanced focal hierarchy | Formulated the visual blueprint: Shoji window left, counter slab center, equipment band, and foreground stools. |\n| **03** | \"Doodle | **FAIL `motor_kinematics`:**
• Straight polygon segments | Applied Wood et al. quadratic Bézier wrist bowing (`Q` curves with $d/L \\in [0.01, 0.04]$ chord normal deflection) and corner vertex overshoots. |\n| **04** | \"Doodle | **FAIL `stroke_hierarchy`:**
• Monotone stroke weights | Enforced 3-tier hierarchy: 4.0px hero contours (shoji/counter), 2.5px props, 1.2px fine interior wood grain and texture hatching. |\n| **05** | \"Doodle | **FAIL `phases_ordered`:**
• Washes painted over ink strokes | Structured DOM into strict phases: Phase 2 watercolor washes rendered *underneath* ink lines with 3–5px deliberate misregistration. |\n| **06** | \"Doodle | **WARN `storytelling_density`:**
• Missing atmospheric life | Added curling coffee steam wisps, falling drip droplets, Kanji calligraphy slats (珈琲, 深煎, 抹茶), and sleeping cat mascot. |\n| **07** | \"Doodle | **WARN `typography_human`:**
• System sans-serif labels | Replaced with organic hand lettering, title banner with coral accent underline, and callout annotations. |\n| **08** | \"Doodle | **PASS (0 violations):**
• AST schema validated
• Archival paper shader rendered | Added zero-JS GPU archival cotton-rag paper texture and micro-scale stroke inking friction shader (`scale=0.8`). |\n\n---\n\n## How to Install & Use with Your AI Agent\n\n### Google Antigravity & Gemini CLI\n\nClone this repository into your agent skills directory or symlink directly:\n\n```bash\n# Clone the skills library\ngit clone https://github.com/surendranb/fun-skills.git ~/skills/fun-skills\n\n# Symlink into your Antigravity / Gemini skills folder\nln -s ~/skills/fun-skills/skills/spatial-architect ~/.gemini/antigravity/skills/spatial-architect\nln -s ~/skills/fun-skills/skills/doodle-architect ~/.gemini/antigravity/skills/doodle-architect\n```\n\n### Claude Code & Claude Desktop\n\nAdd the skills to your project root or Claude config:\n\n```bash\n# Copy into your active workspace .claude/skills directory\nmkdir -p .claude/skills\ncp -r ~/skills/fun-skills/skills/* .claude/skills/\n```\n\n### Cursor, Windsurf, or Any LLM Harness\n\nSimply instruct your coding assistant to follow the prompt rules in [`skills/spatial-architect/SKILL.md`](skills/spatial-architect/SKILL.md) or [`skills/doodle-architect/SKILL.md`](skills/doodle-architect/SKILL.md).\n\n**Prompting Example:**\n> *\"Using the Spatial Architect skill, generate a 1440x900 SVG interior scene for a 1970s Tokyo jazz kissa at 11:00 PM. Follow the 6-phase pipeline, declare datum at y=720, derive all casework from scale 288px=1m, and evaluate the scene with eval_scene.py.\"*\n\n---\n\n## Running the Deterministic Evaluation Suites\n\n```bash\n# 1. Evaluate Spatial SVG Scene (31 Physical AST & Pixel-Tier Checks)\npython3 skills/spatial-architect/scripts/eval_scene.py examples/spatial/01-kyoto-kissaten/kyoto-kissaten.html --render\n\n# 2. Run the 31-Rule Physics Test Suite\npython3 skills/spatial-architect/scripts/test_rules.py\n\n# 3. Evaluate Doodle SVG Sketchnote (Kinematic AST & Palette Compliance)\npython3 skills/doodle-architect/scripts/eval_doodle.py examples/doodle/01-kyoto-kissaten/kyoto-kissaten-doodle.svg\n```\n\n---\n\n## Repository Structure\n\n```\nfun-skills/\n├── README.md # 8-Shot progression grids & eval loop documentation\n├── LICENSE\n├── plugin.json\n├── pyproject.toml\n│\n├── skills/\n│ ├── spatial-architect/ # Realistic 2.5D interior scenes engine\n│ │ ├── SKILL.md # 6-stage physical pipeline prompt\n│ │ ├── references/ # Mathematical laws (F1–F8, L1–L10) & rubric\n│ │ └── scripts/ # 52KB eval_scene.py + 31-rule test suite\n│ │\n│ └── doodle-architect/ # Hand-inked sketchnotes engine\n│ ├── SKILL.md # Motor kinematics & wrist bowing prompt\n│ ├── references/ # Kinematic formulas & watercolor shaders\n│ └── scripts/ # eval_doodle.py AST evaluator\n│\n└── examples/\n ├── spatial/\n │ └── 01-kyoto-kissaten/ # Master Spatial scene, 8 shots, grid & PDF\n │\n └── doodle/\n └── 01-kyoto-kissaten/ # Master Doodle scene, 8 shots, grid & PDF\n```\n\n---\n\n## License\n\n[MIT](LICENSE) © 2026 Surendran Balachandran\n"}