Sketcher
Sketcher turns a spatial brief into a to-scale, code-aware 2D floor plan. It exists because a layperson describing a space ("15 by 10 metre shell, I want a 40-cover restaurant with an open kitchen") can't picture whether it fits, flows, or meets exit rules — Sketcher makes that concrete and drawable.
The engine is schema-first: everything flows through one JSON plan (references/schema.md). Read that file before authoring a plan — the coordinate convention (meters, floor on XZ, Y up, north-up rendering) is the one thing that trips people up. Get the JSON right and rendering is a single script call.
What you can do today (Phase 1)
- Author a plan as JSON matching the schema.
- Render it to 2D SVG — poché walls, real furniture symbols, door swing arcs, windows, room labels with computed areas, a 1-metre scale bar, and a north arrow.
- Lint it against ADA/IBC or India NBC 2016 clearances — exit-door width, egress capacity, room connectivity, furniture overlaps/wall-collisions, and the kitchen work-triangle.
- Load the example-cafe preset (
assets/presets/example-cafe.json) as a starting point for a small F&B space.
Isometric 2.5D, interactive Three.js 3D, and photoreal render-prompt lanes are planned but not built yet — the schema already reserves what they need (wall height, furniture size[2], window sill/head), so they bolt on without breaking existing plans. If the user asks for 3D today, say the 3D lane is a later phase and offer the accurate 2D plan now.
The workflow
Don't jump straight to JSON. A good plan comes from thinking like a designer first — walk the 8-step space-planning method in references/standards.md. In practice:
- Quantify the program with the user in prose — how many covers/desks/people, what activities, what's the real footprint. Pull per-person areas and clearances from
references/standards.md and reality-check that the program even fits before drawing. This is the highest-leverage step; skipping it is how bad layouts happen.
- Author the JSON plan per
references/schema.md. Start walls (the shell), then rooms (zones), then doors/windows, then furniture. For a small F&B space, copy assets/presets/example-cafe.json and edit.
- Lint it:
python3 scripts/lint_plan.py plan.json. Resolve every 🔴 ERROR (they're egress/reference problems). ⚠ WARNings are judgment calls — surface them to the user, don't silently "fix" a spread-out kitchen if the user has a reason.
- Render it:
python3 scripts/render_2d.py plan.json -o plan.svg. Then actually show it — open in a browser or screenshot it — so the user sees the space. Don't just hand over a JSON file.
- Iterate on the user's reaction. It's their space; they drive.
Running the tools
# lint (exit 0 = no errors; prints ERROR/WARN/INFO with sources)
python3 scripts/lint_plan.py plan.json
# render to SVG
python3 scripts/render_2d.py plan.json -o plan.svg
Both are stdlib-only Python 3 — no pip installs. Paths are relative to the skill directory; use absolute paths when running from elsewhere.
Authoring tips that avoid the common traps
- Meters, always. If the user talks in feet, convert on the way in and tell them you did. A plan in feet renders 3.3× too small.
- Walls are the skeleton. Doors and windows anchor to a wall by
id and a fraction t along it, so they ride along if you move the wall. Windows only go on exterior: true walls — the linter enforces it.
use on rooms drives the linter. Prefer the known vocabulary (dining, kitchen, gym, coworking, circulation, toilet, storage, entry) so occupant-load and clearance checks fire correctly.- Unknown furniture types don't crash — they render as a labeled box. But adding a real symbol to
SYMBOLS in scripts/render_2d.py is cheap and makes plans read far better.
- Always emit the disclaimer: clearance checks are planning heuristics, not a building-code certification — a licensed architect must approve the real drawing. Say this whenever you present a plan, especially for assembly-occupancy spaces like restaurants and cafes.
Reference files
references/schema.md — the plan JSON contract. Read before authoring.references/standards.md — clearance numbers (ADA/IBC + India NBC 2016), per-person areas, and the 8-step space-planning method.assets/presets/example-cafe.json — small cafe starter preset (~60 sq m; built to be re-measured and edited for a real space).
1---2name: sketcher3description: Turn room dimensions, a napkin brief, or "help me lay out this space" into an accurate, to-scale 2D floor plan (SVG) with real furniture, doors, windows, and clearance/code sanity checks. Use this whenever the user wants to design, sketch, plan, lay out, or visualize a physical space — a restaurant, cafe, kitchen, gym, coworking floor, office, shop, apartment, or any room — even if they don't say "floor plan": phrases like "how should I arrange", "where should the tables go", "draw my space", "plan the layout", "is this big enough", "will it fit", "design the interior", or giving raw dimensions of a room all trigger it. Also use to sanity-check an existing layout against accessibility and egress clearances (ADA/IBC or India NBC 2016). Produces a JSON plan + rendered SVG you can open in a browser. NOT for pixel-photoreal renders (it writes the prompt for an image AI to do that) and NOT for CAD/DWG output.4---56# Sketcher78Sketcher turns a spatial brief into a **to-scale, code-aware 2D floor plan**. It exists because a layperson describing a space ("15 by 10 metre shell, I want a 40-cover restaurant with an open kitchen") can't picture whether it fits, flows, or meets exit rules — Sketcher makes that concrete and drawable.910The engine is **schema-first**: everything flows through one JSON plan (`references/schema.md`). Read that file before authoring a plan — the coordinate convention (meters, floor on XZ, Y up, north-up rendering) is the one thing that trips people up. Get the JSON right and rendering is a single script call.1112## What you can do today (Phase 1)1314- **Author a plan** as JSON matching the schema.15- **Render it to 2D SVG** — poché walls, real furniture symbols, door swing arcs, windows, room labels with computed areas, a 1-metre scale bar, and a north arrow.16- **Lint it** against ADA/IBC or India NBC 2016 clearances — exit-door width, egress capacity, room connectivity, furniture overlaps/wall-collisions, and the kitchen work-triangle.17- **Load the example-cafe preset** (`assets/presets/example-cafe.json`) as a starting point for a small F&B space.1819Isometric 2.5D, interactive Three.js 3D, and photoreal render-prompt lanes are **planned but not built yet** — the schema already reserves what they need (wall `height`, furniture `size[2]`, window `sill`/`head`), so they bolt on without breaking existing plans. If the user asks for 3D today, say the 3D lane is a later phase and offer the accurate 2D plan now.2021## The workflow2223Don't jump straight to JSON. A good plan comes from thinking like a designer first — walk the **8-step space-planning method** in `references/standards.md`. In practice:24251. **Quantify the program with the user in prose** — how many covers/desks/people, what activities, what's the real footprint. Pull per-person areas and clearances from `references/standards.md` and reality-check that the program even fits *before* drawing. This is the highest-leverage step; skipping it is how bad layouts happen.262. **Author the JSON plan** per `references/schema.md`. Start walls (the shell), then rooms (zones), then doors/windows, then furniture. For a small F&B space, copy `assets/presets/example-cafe.json` and edit.273. **Lint it**: `python3 scripts/lint_plan.py plan.json`. Resolve every 🔴 ERROR (they're egress/reference problems). ⚠ WARNings are judgment calls — surface them to the user, don't silently "fix" a spread-out kitchen if the user has a reason.284. **Render it**: `python3 scripts/render_2d.py plan.json -o plan.svg`. Then actually show it — open in a browser or screenshot it — so the user sees the space. Don't just hand over a JSON file.295. **Iterate** on the user's reaction. It's their space; they drive.3031## Running the tools3233```bash34# lint (exit 0 = no errors; prints ERROR/WARN/INFO with sources)35python3 scripts/lint_plan.py plan.json3637# render to SVG38python3 scripts/render_2d.py plan.json -o plan.svg39```4041Both are stdlib-only Python 3 — no pip installs. Paths are relative to the skill directory; use absolute paths when running from elsewhere.4243## Authoring tips that avoid the common traps4445- **Meters, always.** If the user talks in feet, convert on the way in and tell them you did. A plan in feet renders 3.3× too small.46- **Walls are the skeleton.** Doors and windows anchor to a wall by `id` and a fraction `t` along it, so they ride along if you move the wall. Windows only go on `exterior: true` walls — the linter enforces it.47- **`use` on rooms drives the linter.** Prefer the known vocabulary (`dining`, `kitchen`, `gym`, `coworking`, `circulation`, `toilet`, `storage`, `entry`) so occupant-load and clearance checks fire correctly.48- **Unknown furniture types don't crash** — they render as a labeled box. But adding a real symbol to `SYMBOLS` in `scripts/render_2d.py` is cheap and makes plans read far better.49- **Always emit the disclaimer**: clearance checks are planning heuristics, not a building-code certification — a licensed architect must approve the real drawing. Say this whenever you present a plan, especially for assembly-occupancy spaces like restaurants and cafes.5051## Reference files5253- `references/schema.md` — the plan JSON contract. **Read before authoring.**54- `references/standards.md` — clearance numbers (ADA/IBC + India NBC 2016), per-person areas, and the 8-step space-planning method.55- `assets/presets/example-cafe.json` — small cafe starter preset (~60 sq m; built to be re-measured and edited for a real space).
mohanraj005-dot