Skip to content
Get packs

Guide · Agent skill

How to use the Scientific Schematic Generator agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The scientific-schematics skill gives an AI agent a workflow for drafting scientific diagrams such as flowcharts, neural network architectures and biological pathways from a plain-language description. This guide explains what an agent skill is, how to install this one, how to ask for a diagram, and what to check before using the result. By the end you will be able to generate a PNG draft and know which parts of it you must verify yourself.

Scientific Schematic GeneratorK-Dense-AI/scientific-agent-skills · MIT

Quick start

  1. 1Download and unzip the skill. Download the scientific-schematics ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Set your API key. Export OPENROUTER_API_KEY in your shell so the scripts can call OpenRouter.
  3. 3Describe the diagram. Ask the agent in plain language, naming the components, connections and document type.
  4. 4Run generation. Let the agent run scripts/generate_schematic.py with an output path under figures/ and a --doc-type.
  5. 5Read the review log. Open name_review_log.json and confirm whether the review ran and what score and critique it recorded.
  6. 6Verify every label. Compare the PNG with your source label by label and relationship by relationship before using it.

What an agent skill is and what this one makes the agent do

An agent skill is a folder containing instructions (a SKILL.md file) and optional scripts and reference files that an AI agent can load when a task matches the skill's description. This skill is for generating scientific diagram drafts.

When triggered, the agent writes a prompt from your description, asks an image model (Nano Banana 2, model google/gemini-3.1-flash-image) to draw a PNG, then sends that image to a second model (Gemini 3.7 Flash, google/gemini-3.7-flash) for a quality review. If the review requests improvement and the budget allows, it regenerates with a revised prompt. The skill allows at most two generations.

  • Output is a raster PNG at whatever resolution the image model returns.
  • Each run saves versioned images such as name_v1.png, a copy of the latest successful version, and name_review_log.json with the score and critique.
  • The review is a heuristic score chosen per document type, not a publisher acceptance check.

Install the scientific-schematics skill

Download the skill ZIP from this page and unzip it into one of these locations. A personal install is available to all your projects; a project install applies only to that project.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

  • Personal: unzip into ~/.claude/skills/
  • Project: unzip into .claude/skills/ inside your project folder
  • Claude apps: Settings → Capabilities → Skills, then upload the skill

Set up Python and your OpenRouter API key

The scripts need Python 3.10 or newer and the requests library. Generation requires an OpenRouter API key, which you can create at the OpenRouter keys page. Store it in the OPENROUTER_API_KEY environment variable rather than writing it into prompts or files.

Run the commands from the skill directory so the script paths resolve.

  • Requirements: Python 3.10+, the requests Python library, network access to openrouter.ai, and OPENROUTER_API_KEY.
  • The key is marked optional in the metadata, but image generation cannot run without it.
export OPENROUTER_API_KEY='your_api_key_here'

Trigger the skill with a natural request

You do not need to type a command. Describe the diagram you want, name the document type, and say what the labels and relationships are. The skill's trigger text covers neural network architectures, system diagrams, flowcharts, biological pathways and complex scientific visualizations.

Specific prompts work better. Name the components and the connections between them. Vague requests such as 'make a flowchart' or 'neural network' give the image model too little to work with.

  • Good: 'CONSORT flowchart showing participant flow from screening (n=500) through randomization to final analysis'
  • Good: 'Simplified EGFR → GRB2/SOS → RAS-GTP → RAF → MEK → ERK pathway; label nucleotide exchange and phosphorylation distinctly'
  • Avoid: 'Make a flowchart'

Run the generator with document types and iterations

The generator accepts an output path with -o, a document type with --doc-type, and a maximum number of iterations with --iterations (the maximum is 2). The document type sets the review threshold: journal is 8.5/10, conference and thesis 8.0/10, poster 7.0/10, presentation 6.5/10, and default 7.5/10. Save files under figures/ so they are easy to find.

If a generation is good enough on the first pass, the loop stops and saves the API calls. Otherwise it regenerates until the threshold is met or the iteration limit is reached.

  • -o figures/name.png sets the output path.
  • --doc-type journal, poster, presentation and others choose the threshold.
  • --iterations 2 caps regeneration at two attempts.
python scripts/generate_schematic.py "MAPK signaling pathway from EGFR to gene transcription" -o figures/mapk_pathway.png --doc-type poster

Check every label and relationship before use

Treat each output as an unchecked draft. The review log records the score, the critique and any early-stop reason. If the log shows "score": null with a "review_error", the image was saved but not verified: look at it yourself and diagnose the error before running again. The run prints 'Review unavailable - image kept, quality not verified' in that case.

Image models can misspell labels, invent elements or place arrows incorrectly. Compare the figure against your source, label by label.

  • Check "reviewed" and "final_reviewed" in the log.
  • Verify numbers in trial flows against your records; for CONSORT, reconcile enrollment, allocation, follow-up and analysis counts.
  • Reconsider AI-generated artwork under the target venue's permissions and disclosure policy.

Safety notes: data, network and scripts

This skill runs Python scripts and sends data over the network. Your prompt goes to OpenRouter for generation, and the generated image goes back to OpenRouter for review, both subject to OpenRouter's and the model providers' data policies.

Do not describe unpublished data, patient information or embargoed material in a prompt. The skill is declared to use Bash, Read, Write and Edit tools, so review the scripts and the commands the agent runs before approving them.

  • Runs scripts: yes. Uses network: yes.
  • Keep identifiable or unpublished content out of prompts and images.
  • Output is PNG only; there is no vector, DPI or TIFF/EPS export.

What people use it for

  • A CONSORT-style participant flow for a trial paper, with counts you then reconcile by hand.
  • A neural network architecture sketch for a thesis chapter or slides.
  • A simplified signalling pathway diagram for a poster.
  • A block diagram of a system design for a conference talk.
  • Quick conceptual drafts that a designer or researcher will redraw or edit.

Frequently asked questions

How do I install the scientific-schematics skill?

Download the skill ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps you can instead upload it under Settings → Capabilities → Skills.

Is the scientific schematics skill free?

The skill is published under the MIT licence. Image generation calls OpenRouter with your own API key, so model usage is billed through your OpenRouter account; the source does not list prices, so check OpenRouter's pricing directly.

What do I need to run it?

You need Python 3.10 or newer, the requests Python library, network access to openrouter.ai, and an OpenRouter API key set as OPENROUTER_API_KEY.

Can it make a journal-quality vector figure?

No. The skill returns raster PNG at whatever resolution the image model gives and has no DPI control or vector export. For PDF, EPS or TIFF you would need to convert the PNG yourself and check it at print size.

Why does my log say review unavailable?

The review step may have hit a rate limit, a content filter, or an unexpected reply format. The image is kept but unscored, with score set to null and the reason in review_error. Look at the image yourself and diagnose the error before generating again.

How does this compare with other diagram tools?

This skill suits quick first drafts from text. Tools that produce editable vector diagrams or let you control layout exactly fit final figures better; this source does not benchmark the skill against other tools, so judge by testing on your own material.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you need a fast text-to-PNG first draft of a flowchart, architecture or pathway for a poster, slide or thesis, and you will check every label yourself. Pick a vector diagram editor or a manual workflow if you need journal-ready vector output, precise DPI, or any figure built from unpublished or patient data that cannot leave your machine.

Download the skill