Discourse Graph skills

Agent skills for working with discourse graphs, packaged as a Claude Code plugin marketplace. Each skill is a standard Agent Skill folder, so other agents can use them too. These are research prototypes, so anticipate that they will change frequently.

What's in it

One plugin, discourse-graph (version in plugins/discourse-graph/.claude-plugin/plugin.json):

Skill	Use it to
evidence-candidates	Stage findings from papers as Evidence candidates, each with a verbatim quote, a figure panel and its methods context
figure-extraction	Find a paper's PDF (Zotero → project folder → ask) and crop the figure panel a statement reports, recording which file and method it came from

Both skills are tool-agnostic. SKILL.md describes each step, and the references/ files list ways to do each step with Zotero, PyMuPDF, poppler, PubMed Central, Roam and Obsidian.

Install

In Claude Code:

/plugin marketplace add DiscourseGraphs/discourse-graph-skills
/plugin install discourse-graph@discourse-graph-skills

The skills then show up as discourse-graph:evidence-candidates and discourse-graph:figure-extraction. If you have no GitHub SSH key, set CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1 before adding the marketplace.

To get updates, run /plugin marketplace update discourse-graph-skills, or turn on auto-update for this marketplace under Marketplaces in /plugin.

The quote checker needs Python 3 with PyMuPDF (pip install pymupdf).

For other agents (Codex, Cursor, Gemini CLI and others), copy a folder from plugins/discourse-graph/skills/ into that agent's skills directory.

Editing

• Edit here, never in an installed copy.
• Bump version in plugin.json for each release. People who installed the plugin get changes only when the version changes.
• Remove personal copies. If you also have a personal skill with the same name in ~/.claude/skills/, both can trigger. Remove the personal copy once you use the plugin.
• Test changes first. Run the scenarios in tests/ with a fresh agent and compare against the recorded results. After changing check_quote.py, run python3 tests/test_check_quote.py.
• Validate: claude plugin validate .

Credits

• evidence-candidates is generalized from the Akamatsu lab's evidence-candidates skill.
• The verbatim-quote rule, the evidence bundle, and the figure-capture methods and PubMed Central route in figure-extraction are adapted from the living-synthesis pipeline (Joel Chan's team, with work by Karola Kirsanow, building on Jay Patel's living-synthesis vault).

License

• The skills, references and documentation are licensed under CC BY 4.0.
• The code (*.py files and the code snippets in the references) is licensed under the MIT License.

Copyright (c) 2026 Discourse Graphs Project contributors.