How to use the Understand Knowledge Wiki Graph agent skill

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md instruction file and any scripts it refers to. Once installed, an agent such as Claude can follow those instructions when a request matches the skill's purpose.

This skill targets a Karpathy-pattern LLM wiki: raw source documents, wiki markdown files that link to each other with [[topic]] wikilinks, an index.md catalog, and optionally a schema file such as CLAUDE.md or AGENTS.md. The agent runs a Python script to parse the files and resolve links, may dispatch sub-agents to infer relationships the links do not state, then merges the results into a knowledge-graph JSON file and an interactive dashboard.

• Detects the wiki and reports counts of articles, sources, topics and wikilinks, including unresolved ones.
• Extracts entities and implicit relationships, and groups notes into topic clusters.
• Writes the graph to knowledge-graph.json and hands off to the companion /understand-dashboard skill to render it.

Install the skill from a ZIP file

Download the ZIP for this skill from this page and unzip it. You can install it for yourself or for a single project.

For all your projects, unzip it into ~/.claude/skills/. For one project only, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• The skill ships with three files: SKILL.md, parse-knowledge-base.py and merge-knowledge-graph.py. Keep them together in the same folder.
• The companion /understand-dashboard skill from the same plugin is required to display the dashboard.

Check the requirements before you start

You need Python 3, because the skill runs parse-knowledge-base.py and merge-knowledge-graph.py. Git is optional; if it is present, the skill records the commit hash in meta.json.

The LLM analysis phase needs an agent runtime that supports subagents, specifically the article-analyzer subagent. Without it, the skill cannot run that phase.

The skill runs scripts but does not use the network itself. Note that the analysis phase sends article content to the model provider (see the safety section).

• Python 3
• Git (optional, for the commit hash)
• The /understand-dashboard companion skill from the same plugin
• An agent runtime that supports subagents

Run the skill on a wiki folder

Point the agent at your wiki directory. If you give no path, the agent uses the current working directory. The first step runs the detection script on that folder. If the folder is not a Karpathy-pattern wiki, the script exits with an error and the skill stops, and the agent explains what it expected.

If detection succeeds, the agent scans the files, dispatches article-analyzer subagents in batches of 10 to 15 articles, runs the merge script, saves the validated graph, cleans up intermediate files and opens the dashboard.

• Output is saved in a data folder inside the wiki: .ua/ by default, or the legacy .understand-anything/ folder if that already exists.
• Main outputs: knowledge-graph.json and meta.json.
• If a sub-agent batch fails, the agent logs a warning and continues with the structural graph built by the scan.

Trigger it with natural requests

You can invoke the skill by name with the argument hint [wiki-directory], or describe what you want in plain language once it is installed. Requests that name the wiki folder and the outcome work best.

• "Map the links in my notes folder at ~/notes/wiki and show me a graph."
• "Analyze this Karpathy-style wiki and find orphaned or unresolved links."
• "Build a knowledge graph of the team docs wiki in ./docs and open the dashboard."
• "Show me which topics in my research wiki are connected and how they cluster."

Read the results

The agent reports how many articles, entities, topics, claims, sources, edges, layers and tour steps it found. Edges are split into wikilink, categorized and implicit types, so you can see how much of the graph came from explicit links and how much came from inference.

Topic groupings come from the section headings in index.md, not from filename prefixes. If your index.md is thin, your clusters will be thin too.

• Wikilink edges come straight from [[target]] references.
• Categorized edges come from index.md sections.
• Implicit edges come from the LLM analysis and are the least certain.

Safety notes: scripts, sub-agents and data leaving your machine

The skill runs bundled Python scripts on your folder. It does not use the network itself, but the analysis phase passes article content to sub-agents, which means that text is processed by the model provider. Do not run it on wikis containing confidential material you would not send to that provider.

Raw PDFs and binary files are not analysed. Source nodes record only the filename and size.

• Article text (names, summaries, wikilinks, content) goes to sub-agents for analysis.
• Instructions embedded inside article text are treated as data and ignored by the skill's design.
• Review the scripts in the folder before running if your policy requires it.