AgentFigureGallery is a drop-in scientific plotting skill for Claude Code, Codex, Cursor, and other coding agents.
It turns real visual references plus human like/reject feedback into action-ready plotting guidance before code is written.
Clone once, run one command, and your coding agent gets a human-curated scientific figure gallery before writing plotting code. Codex, Claude Code, and Cursor-compatible installs are supported.
AgentFigureGallery helps coding agents stop guessing what a publication figure should look like. The agent queries visual references first, the human marks taste preferences in a browser gallery, and the selected references are exported as an action bundle before plotting code is written.
With the 16k+ full-public reference pool, that browser gallery becomes a routine taste-tuning loop: launch it often, like/reject/select references, and gradually adapt the skill to your personal or lab-specific figure preferences.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
First-time install walkthrough for Claude Code, Codex CLI, and ChatGPT.
After installing the Codex skill, Codex can discover AgentFigureGallery as a local skill.
Then ask your coding agent to run a plot-type smoke test:
Use AgentFigureGallery to test your installed plotting skill. Generate one Nature-style example for each supported plot type, then export PNG/PDF/SVG and a combined preview.
The result should look like this: one Nature-style smoke example for every supported plot type.
See examples/plot_type_examples/ for the runnable script, source data, and PNG/PDF/SVG outputs.
Launch the browser gallery UI after install to refine your personal gallery:
agentfiguregallery gallery --plot-type embedding_plot --limit 50 --serve
# Then open http://127.0.0.1:8765/
Use it routinely to browse the reference pool and record like/reject/select feedback; those preferences become reusable taste memory for future agent plotting tasks. After your agent expands the gallery, or after you drop new visible references into a local pack, relaunch the gallery and keep refining the same preference memory.
To reopen the frontend later without creating a new reference session:
Codex installs to ~/.codex/skills, Claude Code installs to ~/.claude/skills, Cursor-compatible installs to ~/.cursor/skills, and Cursor Project Rules install to .cursor/rules/agent-figure-gallery.mdc. See docs/AGENT_QUICKSTART.md and examples/agent_prompt.md.
Use the browser gallery to generate candidates by plot type from the 16,341-candidate full-public KB, remove bad references globally, keep type-specific preferences, and export selected references for the agent that will write the final plotting code. Every like/reject becomes reusable taste memory, so the skill gets closer to your visual taste as humans and agents keep using it.
AgentFigureGallery is designed to grow after install. You can ask an agent to follow the expansion contract, or add a small local reference pack yourself, then use the browser gallery to fold the new material into your taste memory.
Tell your coding agent:
Read ExtendAgent/README.md, then expand AgentFigureGallery for <plot type or style>. Discover high-quality public scientific plotting sources, render every useful reference as a visible preview, preserve stable candidate IDs and source license metadata, rebuild the candidate index, and report candidate counts plus private-path scan results.
For manual expansion, use the same contract:
Add only references that have a visible preview PNG; screenshots or scripts without previews cannot enter the human selection loop.
Give every reference a stable candidate_id, plot_type, preview path, source repository or file metadata, and license/source attribution when available.
Keep large preview packs and raw upstream repositories out of Git; publish or store them as packs and update the manifest when they should be shared.
Preserve existing preference memory in data/reference_global_preferences.json and outputs/reference_sessions/**/preferences.json.
Refresh the candidate index, run agentfiguregallery doctor, then launch agentfiguregallery gallery --plot-type <plot_type> --limit 50 --serve to inspect and refine the new candidates.
See ExtendAgent/README.md for the maintainer-oriented expansion rules and quality gates.
What Is Inside
16,341 full-public visual candidates across 10 scientific plot types.
Routine browser-gallery feedback that adapts the skill to personal or lab-specific figure preferences.
Glike-curated minimal pack committed for instant smoke tests.
Codex-equipped plot-type smoke examples with PNG/PDF/SVG outputs.
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
