Installable skillMIT

canvas-layout: Re-Layout Canvas Nodes

Re-layout existing Obsidian Canvas nodes using 6 spatial algorithms: grid (galleries, mood boards), dagre (flowcharts, org charts), radial (mind maps), force-directed (knowledge graphs), linear (timelines), and auto-detect. Preserves group membership, snaps to 20px grid, refits zones around conte...

BY AgriciDanielUPDATED Oct 27, 2026
INTENDED FOR◈OpenClawInstall in a listed agent.

instructions.md

---
name: canvas-layout
description: >
  Re-layout existing Obsidian Canvas nodes using 6 spatial algorithms: grid
  (galleries, mood boards), dagre (flowcharts, org charts), radial (mind maps),
  force-directed (knowledge graphs), linear (timelines), and auto-detect.
  Preserves group membership, snaps to 20px grid, refits zones around content.
  Triggers on: canvas layout, re-layout canvas, arrange canvas, auto-layout,
  organize canvas, fix canvas layout, canvas grid, canvas dagre.
user-invocable: false
---

# canvas-layout: Re-Layout Canvas Nodes

Read `../canvas/references/layout-algorithms.md` for algorithm details and selection guide.
Read `../canvas/references/performance-guide.md` for node limits.

---

## Workflow

1. **Identify target canvas**: Use the active canvas or ask which one (`/canvas list` to show options).
2. **Select algorithm**: Use the user-specified algorithm, or run `auto` detection.
3. **Confirm before applying**: Show what will happen:
   "Will apply [algorithm] layout to [canvas] ([N] nodes, [M] groups). Create backup? [Y/n]"
4. **Run the layout script**:
   ```bash
   python3 scripts/canvas_layout.py [canvas_path] [algorithm] [options]
   ```
5. **Report results**: "Moved [N] of [M] nodes. [G] groups refitted. Backup at [path].bak"
6. **Validate**: Run `python3 scripts/canvas_validate.py [canvas_path]` to confirm valid output.

---

## Algorithm Selection

When the user says `/canvas layout auto` or doesn't specify an algorithm, use the auto-detection in the script. It analyzes edge density, node types, and connection patterns.

For explicit requests, map user intent to algorithm:

| User says | Algorithm | Options |
|-----------|-----------|---------|
| "organize these images" | grid | `--sort-by type` |
| "make a grid" | grid | |
| "flowchart layout" | dagre | `--direction TB` |
| "left to right flow" | dagre | `--direction LR` |
| "mind map layout" | radial | |
| "expand from [node]" | radial | `--center [node-id]` |
| "untangle this" | force | |
| "spread out the nodes" | force | |
| "make a timeline" | linear | `--axis horizontal` |
| "vertical sequence" | linear | `--axis vertical` |
| "auto-layout" | auto | |
| "fix the layout" | auto | |

---

## Options Forwarding

Pass algorithm-specific options to the script:

- **grid**: `--columns N` (override auto-detection), `--sort-by type|size`
- **dagre**: `--direction TB|LR|BT|RL`
- **radial**: `--center node-id` (override auto-detection of hub node)
- **force**: `--iterations N` (default 100, reduce to 50 for 50+ nodes)
- **linear**: `--axis horizontal|vertical`

Add `--dry-run` to preview without writing.

---

## For Complex Layouts (30+ Nodes)

When a canvas has more than 30 nodes, dispatch the `canvas-layout` agent instead of running the script directly. The agent can:

- Analyze the canvas content to choose sub-groups for different algorithms
- Split large canvases into zone-by-zone layouts
- Handle edge crossing minimization that the basic dagre doesn't cover
- Refine positions after the algorithm runs for better visual balance

---

## Edge Cases

- **Empty canvas**: Report "No nodes to layout" and skip.
- **Single node**: Center it at (0, 0) and skip.
- **No edges + dagre requested**: Warn "dagre works best with edges. Using grid instead." Fall back to grid.
- **100+ nodes**: Warn about performance. Suggest splitting into sub-canvases.
- **Backup conflict**: If `.bak` already exists, use `.bak2`, `.bak3`, etc.
COMPATIBILITY & REFERENCES

Platform guidance

Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.