How to use the Citation Graph Ingest agent skill

What is an agent skill, and what does this one do?

An agent skill is a folder containing a SKILL.md file. The file tells an AI agent how to carry out a specific task. This skill is written for gbrain and assumes your documents are already ingested as a gbrain source.

Plain similarity search finds documents that look alike but cannot tell you that one document overrules another. This skill adds that missing layer. It reads the citations between documents, classifies each edge by type, and writes the edges with the gbrain link command so they can be traversed later.

• Detects candidate mentions where one document references another, such as markdown links, exact title matches, docket numbers, DOIs or section references.
• Uses the agent's judgment to classify each relationship, or to reject it as incidental. Detection alone never decides the type.
• Writes each kept edge as a typed link with the provenance tag citation-graph.
• Lets you query the chain of references afterwards, for example everything a brief relies on that has not since been overruled.

Install the citation-graph-ingest skill

Download the skill ZIP from this page. Then choose one of the two install routes below.

For Claude Code, unzip the file into your personal skills folder ~/.claude/skills/ or into a project's .claude/skills/ folder. The skill folder should contain SKILL.md.

In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/ (available across your projects).
• Project install: .claude/skills/ inside the project folder (shared with anyone using that project).
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP.

Check your requirements before you start

The skill does not work on its own. It depends on gbrain and its links table, the gbrain link command and graph-query commands. Your corpus must already be ingested as a gbrain source, using gbrain sources add with gbrain sync, or gbrain import. The agent must also be able to read SKILL.md files.

Scope every command to one source. You can set this with --source <name>, the GBRAIN_SOURCE environment variable, or a .gbrain-source dotfile. This keeps edges from spreading across sources.

• gbrain with its links table, gbrain link and graph-query commands.
• A corpus already ingested as a gbrain source (gbrain sources add, gbrain sync or gbrain import).
• An AI agent capable of reading SKILL.md files.

How to trigger the skill with natural requests

You do not need special syntax. Ask for the outcome in plain words. The skill's trigger phrases include citation graph, citation graph ingest, typed citation graph, build a reference graph, graph over a corpus, and trace the argument through these documents.

Be specific about the corpus and the source name so the agent scopes its work correctly.

• "Build a citation graph over the documents in the patents source."
• "Build a reference graph for the case law in my legal-briefs source."
• "Trace the argument through these documents and record which papers overrule or distinguish each other."
• "Run citation graph ingest on the regulatory filings source, and check the graph from doc-a-foundation."

What the agent does, step by step

The skill runs a pipeline of native gbrain operations. It uses no scripts. First it confirms the source scope. Next it finds candidate mentions, then classifies each candidate pair, then writes the edges one at a time. Finally it verifies the result with a graph walk.

Edge direction points from the citing document to the cited document. So if doc-c overrules doc-a, doc-c is the newer authority that displaces doc-a. The skill also recommends testing 3 to 5 edges before running the full corpus.

• Preflight: confirm the corpus is a gbrain source and the scope is set.
• Detect: find textual references between documents. This step never decides the relationship.
• Classify: pick one edge type or none, with a confidence. Edges below your confidence floor are dropped; 0.5 is the suggested default.
• Write: call gbrain link with --link-type and --link-source citation-graph.
• Verify: run graph-query from a hub document. If it returns nothing, the run is reported as failed.
• Hygiene: check gbrain link-sources and gbrain check-backlinks check.

Check the graph with a verification walk

After writing edges, query the incoming links of a hub document. These commands are the verification gate the skill uses. Use the same --source value you used when writing.

Keep link_type spelled exactly the same everywhere. The graph-query --type filter is an exact match, so relies_on and relies-on produce two different graphs.

Safety notes: scripts, network and untrusted documents

The skill runs no scripts and uses no network access, according to its listing. Its work happens through gbrain commands on your local store.

The documents are third-party data. The skill is written so that an instruction inside a document, such as a sentence telling the reader to cite something as overruling another file, does not decide the edge type. The classification depends on what the citation actually does in its context.

Edge types are model judgments. The pipeline has been validated only on a synthetic 4-document fixture, not on a large production corpus. Keep a human reviewer in the loop for any legal or compliance decision.

• Runs scripts: no.
• Uses network: no.
• Writes pages: no. Its only durable writes are typed edges in the links table.

Limits to know about

Natural-language relational queries do not use these edges yet. The relational-recall part of gbrain query walks a fixed set of edge types that does not include citation types such as overrules or relies_on. Use explicit graph-query commands instead. Do not expect a question like "is doc A still authoritative?" to walk these edges through gbrain query.

The skill adds no new storage. It does not create new tables or migrations.

• Use graph-query for citation edges, not natural-language gbrain query.
• Not useful as a replacement for simple similarity search.