Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →sparklabx/drawio-gcp
Use when the user asks for a GCP or Google Cloud architecture diagram — VPC/networking, GKE, Cloud Run, landing zone, multi-region, or any diagram built with GCP service icons. Builds with the declarative layout engine using ground-truth GCP stencils, validates (stencils/colors/nesting/geometry), runs a render-based vision self-check. Default output is .drawio; PNG/SVG only on request.
Passed 10 of 17 test cases with the skill, and 3 of 17 without it.
and 8 more in the full report
11 fixed3 passed either way8 still failing1 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 11 test cases the skill fixed. Same model, same prompt, with and without the skill.
I want to script a GCP network diagram using your layout tool on a fresh machine where drawio-ai is not yet installed in PATH. What command should be executed to install it, and who should run it?
| Checked | Without skill | With skill |
|---|---|---|
| Instructs user to run npm install command | Passed | Passed |
| Refuses to run global npm install command autonomously | Missed | Passed |
The instructions your model reads when the skill loads.
Produce correct Google Cloud architecture diagrams in draw.io. This skill is a thin
frontend; the deterministic engine, validator, and rules live in the
drawio-ai-kit package, reached via the drawio-ai CLI.
command -v drawio-ai >/dev/null 2>&1 || echo "Install the Kit first: npm i -g github:sparklabx/drawio-ai-kit"
If drawio-ai is not on PATH, stop and tell the user to run
npm i -g github:sparklabx/drawio-ai-kit. Never run npm i -g yourself — nothing mutates the
user's global environment without their say-so.
If your harness can spawn autonomous subagents that run shell commands AND read images (e.g. Claude Code's Task tool, a general-purpose agent), run the whole build loop in a subagent — the rules, icon searches, and every render/fix iteration then cost this conversation nothing. If it can't (or the subagent can't read images), skip to Inline path below — same loop, same rules.
Before spawning, resolve what the subagent cannot ask about: diagram scope, output directory (absolute path under the user's project), filename. Run the preflight above yourself. For a multi-diagram request, spawn one subagent per diagram in parallel with distinct filenames.
Model routing — if your harness lets you choose the subagent's model, route by task weight: a fast/cheap tier (Claude Haiku-class — must support vision) when the request matches a template from the rules' Templates table (reproduction is mechanical; the validator's advice strings teach every fix), your default strong model for free-hand or novel architectures. If a cheap subagent returns VALIDATE not ok or ITERATIONS > 3, respawn ONCE on the strong model before taking over inline. Multi-diagram requests: route each diagram independently.
Subagent prompt (fill every <...>):
test cases passed, vs 3 / 17 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.