// Create Excalidraw diagram JSON files that make visual arguments. Use when the user wants to visualize workflows, architectures, or concepts.
Triage one tt-forge-models training test failing with a bfloat16 dtype-mismatch RuntimeError (e.g. "mat1 and mat2 must have the same dtype, but got Float and BFloat16", "'<op>' not implemented for 'BFloat16'"). For cross-dtype operands, attempts a minimal loader fix propagating `dtype_override` into the offending tensor constructor, then re-runs CPU + pytest and updates the YAML (passing -> EXPECTED_PASSING; new failure -> KNOWN_FAILURE_XFAIL). For op-not-implemented (no PyTorch kernel), goes straight to KNOWN_FAILURE_XFAIL with the verbatim error. Updates every training entry sharing the affected loader. Never edits inference YAML or `dynamic_loader.py`.
Triage one tt-forge-models training test stuck at FAILED_FE_COMPILATION with reason "tt-forge-models doesn't implement unpack_forward_output for this model." Inspects the model's forward output, registers a handler or writes a per-loader override, and updates the YAML.
Analyze CI benchmark workflow runs from GitHub Actions for the tt-xla project. Produces a markdown report covering failed jobs (with root-cause error extraction via logs and Glean), successful model performance metrics (samples/sec, TTFT, device perf), perf regressions/improvements vs previous nightly, and the full dependency commit chain (tt-xla, tt-mlir, tt-metal). Use this skill whenever the user wants to analyze a CI run, review nightly benchmark results, investigate CI failures, check benchmark performance from a workflow run, or asks about "latest nightly" results. Also trigger when the user pastes a GitHub Actions run URL or mentions a run ID in the context of performance analysis, or asks about perf regressions.
Use when auditing a TTNN model's IR for missed op fusion opportunities — both direct TTNN fusions (a fused ttnn op already exists) and theoretical fusions (the pattern is a single kernel in torch/triton/cuda)
Analyze a GitHub Actions run and summarize failures
Analyzes, debugs and proposes fixes for graph breaks in PyTorch/XLA model compilation. Use when a model generates more graphs than expected during compilation, the user mentions "graph break", or when debugging excessive graph generation in tt-xla pipelines.
| name | excalidraw-diagram |
| description | Create Excalidraw diagram JSON files that make visual arguments. Use when the user wants to visualize workflows, architectures, or concepts. |
Generate .excalidraw JSON files that argue visually, not just display information.
All colors and brand-specific styles live in one file: references/color-palette.md. Read it before generating any diagram and use it as the single source of truth for all color choices — shape fills, strokes, text colors, evidence artifact backgrounds, everything.
To make this skill produce diagrams in your own brand style, edit color-palette.md. Everything else in this file is universal design methodology and Excalidraw best practices.
Diagrams should ARGUE, not DISPLAY.
A diagram isn't formatted text. It's a visual argument that shows relationships, causality, and flow that words alone can't express. The shape should BE the meaning.
The Isomorphism Test: If you removed all text, would the structure alone communicate the concept? If not, redesign.
The Education Test: Could someone learn something concrete from this diagram, or does it just label boxes? A good diagram teaches—it shows actual formats, real event names, concrete examples.
Before designing, determine what level of detail this diagram needs:
Use abstract shapes when:
Use concrete examples when:
For technical diagrams, you MUST include evidence artifacts (see below).
Before drawing anything technical, research the actual specifications.
If you're diagramming a protocol, API, or framework:
Bad: "Protocol" → "Frontend" Good: "AG-UI streams events (RUN_STARTED, STATE_DELTA, A2UI_UPDATE)" → "CopilotKit renders via createA2UIMessageRenderer()"
Research makes diagrams accurate AND educational.
Evidence artifacts are concrete examples that prove your diagram is accurate and help viewers learn. Include them in technical diagrams.
Types of evidence artifacts (choose what's relevant to your diagram):
| Artifact Type | When to Use | How to Render |
|---|---|---|
| Code snippets | APIs, integrations, implementation details | Dark rectangle + syntax-colored text (see color palette for evidence artifact colors) |
| Data/JSON examples | Data formats, schemas, payloads | Dark rectangle + colored text (see color palette) |
| Event/step sequences | Protocols, workflows, lifecycles | Timeline pattern (line + dots + labels) |
| UI mockups | Showing actual output/results | Nested rectangles mimicking real UI |
| Real input content | Showing what goes IN to a system | Rectangle with sample content visible |
| API/method names | Real function calls, endpoints | Use actual names from docs, not placeholders |
Example: For a diagram about a streaming protocol, you might show:
Example: For a diagram about a data transformation pipeline:
The key principle: show what things actually look like, not just what they're called.
Comprehensive diagrams operate at multiple zoom levels simultaneously. Think of it like a map that shows both the country borders AND the street names.
A simplified overview showing the full pipeline or process at a glance. Often placed at the top or bottom of the diagram.
Example: Input → Processing → Output or Client → Server → Database
Labeled regions that group related components. These create visual "rooms" that help viewers understand what belongs together.
Example: Grouping by responsibility (Backend / Frontend), by phase (Setup / Execution / Cleanup), or by team (User / System / External)
Evidence artifacts, code snippets, and concrete examples within each section. This is where the educational value lives.
Example: Inside a "Backend" section, you might show the actual API response format, not just a box labeled "API Response"
For comprehensive diagrams, aim to include all three levels. The summary gives context, the sections organize, and the details teach.
| Bad (Displaying) | Good (Arguing) |
|---|---|
| 5 equal boxes with labels | Each concept has a shape that mirrors its behavior |
| Card grid layout | Visual structure matches conceptual structure |
| Icons decorating text | Shapes that ARE the meaning |
| Same container for everything | Distinct visual vocabulary per concept |
| Everything in a box | Free-floating text with selective containers |
| Simple Diagram | Comprehensive Diagram |
|---|---|
| Generic labels: "Input" → "Process" → "Output" | Specific: shows what the input/output actually looks like |
| Named boxes: "API", "Database", "Client" | Named boxes + examples of actual requests/responses |
| "Events" or "Messages" label | Timeline with real event/message names from the spec |
| "UI" or "Dashboard" rectangle | Mockup showing actual UI elements and content |
| ~30 seconds to explain | ~2-3 minutes of teaching content |
| Viewer learns the structure | Viewer learns the structure AND the details |
Simple diagrams are fine for abstract concepts, quick overviews, or when the audience already knows the details. Comprehensive diagrams are needed for technical architectures, tutorials, educational content, or when you want the diagram itself to teach.
Not every piece of text needs a shape around it. Default to free-floating text. Add containers only when they serve a purpose.
| Use a Container When... | Use Free-Floating Text When... |
|---|---|
| It's the focal point of a section | It's a label or description |
| It needs visual grouping with other elements | It's supporting detail or metadata |
| Arrows need to connect to it | It describes something nearby |
| The shape itself carries meaning (decision diamond, etc.) | Typography alone creates sufficient hierarchy |
| It represents a distinct "thing" in the system | It's a section title, subtitle, or annotation |
Typography as hierarchy: Use font size, weight, and color to create visual hierarchy without boxes. A 28px title doesn't need a rectangle around it.
The container test: For each boxed element, ask "Would this work as free-floating text?" If yes, remove the container.
Before anything else, determine if this needs to be:
If comprehensive: Do research first. Look up actual specs, formats, event names, APIs.
Read the content. For each concept, ask:
For each concept, find the visual pattern that mirrors its behavior:
| If the concept... | Use this pattern |
|---|---|
| Spawns multiple outputs | Fan-out (radial arrows from center) |
| Combines inputs into one | Convergence (funnel, arrows merging) |
| Has hierarchy/nesting | Tree (lines + free-floating text) |
| Is a sequence of steps | Timeline (line + dots + free-floating labels) |
| Loops or improves continuously | Spiral/Cycle (arrow returning to start) |
| Is an abstract state or context | Cloud (overlapping ellipses) |
| Transforms input to output | Assembly line (before → process → after) |
| Compares two things | Side-by-side (parallel with contrast) |
| Separates into phases | Gap/Break (visual separation between sections) |
For multi-concept diagrams: each major concept must use a different visual pattern. No uniform cards or grids.
Before JSON, mentally trace how the eye moves through the diagram. There should be a clear visual story.
Only now create the Excalidraw elements. See below for how to handle large diagrams.
For comprehensive or technical diagrams, you MUST build the JSON one section at a time. Do NOT attempt to generate the entire file in a single pass. This is a hard constraint — Claude Code has a ~32,000 token output limit per response, and a comprehensive diagram easily exceeds that in one shot. Even if it didn't, generating everything at once leads to worse quality. Section-by-section is better in every way.
Phase 1: Build each section
type, version, appState, files) and the first section of elements."trigger_rect", "arrow_fan_left") so cross-section references are readable.boundElements array at the same time.Phase 2: Review the whole
After all sections are in place, read through the complete JSON and check:
Fix any alignment or binding issues.
Plan your sections around natural visual groupings from the diagram plan. A typical large diagram might split into:
Each section should be independently understandable: its elements, internal arrows, and any cross-references to adjacent sections.
Central element with arrows radiating to multiple targets. Use for: sources, PRDs, root causes, central hubs.
○
↗
□ → ○
↘
○
Multiple inputs merging through arrows to single output. Use for: aggregation, funnels, synthesis.
○ ↘
○ → □
○ ↗
Parent-child branching with connecting lines and free-floating text (no boxes needed). Use for: file systems, org charts, taxonomies.
label
├── label
│ ├── label
│ └── label
└── label
Use line elements for the trunk and branches, free-floating text for labels.
Elements in sequence with arrow returning to start. Use for: feedback loops, iterative processes, evolution.
□ → □
↑ ↓
□ ← □
Overlapping ellipses with varied sizes. Use for: context, memory, conversations, mental states.
Input → Process Box → Output with clear before/after. Use for: transformations, processing, conversion.
○○○ → [PROCESS] → □□□
chaos order
Two parallel structures with visual contrast. Use for: before/after, options, trade-offs.
Visual whitespace or barrier between sections. Use for: phase changes, context resets, boundaries.
Use lines (type: line, not arrows) as primary structural elements instead of boxes:
Timeline: Tree:
●─── Label 1 │
│ ├── item
●─── Label 2 │ ├── sub
│ │ └── sub
●─── Label 3 └── item
Lines + free-floating text often creates a cleaner result than boxes + contained text.
Choose shape based on what it represents—or use no shape at all:
| Concept Type | Shape | Why |
|---|---|---|
| Labels, descriptions, details | none (free-floating text) | Typography creates hierarchy |
| Section titles, annotations | none (free-floating text) | Font size/weight is enough |
| Markers on a timeline | small ellipse (10-20px) | Visual anchor, not container |
| Start, trigger, input | ellipse | Soft, origin-like |
| End, output, result | ellipse | Completion, destination |
| Decision, condition | diamond | Classic decision symbol |
| Process, action, step | rectangle | Contained action |
| Abstract state, context | overlapping ellipse | Fuzzy, cloud-like |
| Hierarchy node | lines + text (no boxes) | Structure through lines |
Rule: Default to no container. Add shapes only when they carry meaning. Aim for <30% of text elements to be inside containers.
Colors encode information, not decoration. Every color choice should come from references/color-palette.md — the semantic shape colors, text hierarchy colors, and evidence artifact colors are all defined there.
Key principles:
Do not invent new colors. If a concept doesn't fit an existing semantic category, use Primary/Neutral or Secondary.
For clean, professional diagrams:
roughness: 0 — Clean, crisp edges. Use for modern/technical diagrams.roughness: 1 — Hand-drawn, organic feel. Use for brainstorming/informal diagrams.Default to 0 for most professional use cases.
strokeWidth: 1 — Thin, elegant. Good for lines, dividers, subtle connections.strokeWidth: 2 — Standard. Good for shapes and primary arrows.strokeWidth: 3 — Bold. Use sparingly for emphasis (main flow line, key connections).Always use opacity: 100 for all elements. Use color, size, and stroke width to create hierarchy instead of transparency.
Instead of full shapes, use small dots (10-20px ellipses) as:
The most important element has the most empty space around it (200px+).
Guide the eye: typically left→right or top→bottom for sequences, radial for hub-and-spoke.
Position alone doesn't show relationships. If A relates to B, there must be an arrow.
CRITICAL: The JSON text property contains ONLY readable words.
{
"id": "myElement1",
"text": "Start",
"originalText": "Start"
}
Settings: fontSize: 16, fontFamily: 3, textAlign: "center", verticalAlign: "middle"
{
"type": "excalidraw",
"version": 2,
"source": "https://excalidraw.com",
"elements": [...],
"appState": {
"viewBackgroundColor": "#ffffff",
"gridSize": 20
},
"files": {}
}
See references/element-templates.md for copy-paste JSON templates for each element type (text, line, dot, rectangle, arrow). Pull colors from references/color-palette.md based on each element's semantic purpose.
text contains only readable wordsfontFamily: 3roughness: 0 for clean/modern (unless hand-drawn style requested)opacity: 100 for all elements (no transparency)