Build a reviewable knowledge graph with ak:graphify

Use ak:graphify to turn code, documents, papers, images, or local media into a queryable relationship graph. It writes reusable visualization, report, JSON, and cache artifacts so you can explore structure before planning or editing.

Choose ak:graphify for relationship exploration

Use ak:graphify when

• You need a high-level map of an unfamiliar codebase or document collection.
• You want to explore dependencies, connected concepts, central nodes, or paths.
• You need persistent graph artifacts that can be queried through an optional MCP server.

Choose another workflow when

• You need symbol-level go-to-definition or find-usages behavior. Use ak:gkg.
• You only need a bounded file map. Use ak:scout.
• You need raw context packaging instead of relationship extraction.
• You cannot send documents or images to the configured model provider and cannot exclude them from the input.

Prepare the input and runtime

Install Python 3.10 or newer and the separately distributed graphifyy package. Choose the smallest input path, review ignored and generated content, estimate disk and provider use, and decide whether an ongoing watch process is allowed.

Run the Skill

/ak:graphify ./src --report "Map module dependencies, keep EXTRACTED and INFERRED edges separate, exclude secrets and generated files, and do not start a watch process or MCP server."

/ak:graphify ./src --report "Map module dependencies, keep EXTRACTED and INFERRED edges separate, exclude secrets and generated files, and do not start a watch process or MCP server."

$ak:graphify ./src --report "Map module dependencies, keep EXTRACTED and INFERRED edges separate, exclude secrets and generated files, and do not start a watch process or MCP server."

Understand the observable stages

1. Resolve scope: inventory selected files, exclusions, media types, and repository instructions before processing.
2. Extract structure: parse supported code with tree-sitter and tag direct imports, calls, or inheritance as EXTRACTED.
3. Process other media: transcribe supported audio or video locally when the needed components exist.
4. Extract semantics: send selected documents or images to the configured model provider and tag derived relationships as INFERRED or AMBIGUOUS.
5. Assemble artifacts: write the graph JSON, interactive HTML, report, and incremental cache beneath graphify-out/.
6. Review or serve: inspect surprising paths and confidence tags; start an MCP server only when separately authorized.

Keep privacy and graph claims bounded

• Do not include secret stores, credentials, private customer material, build output, or unrelated archives merely because they are under the input path.
• An EXTRACTED edge is parser evidence, not proof of runtime behavior or architectural intent. An INFERRED edge is a model interpretation and needs source review. AMBIGUOUS edges must remain visibly unresolved.
• Installing graphifyy or extras changes the Python environment. The upstream graphify install command is not needed to install the AgentKit Skill.
• Starting watch or MCP modes does not authorize persistent background service, network exposure, or edits to client configuration.
• Review existing graphify-out/ content before overwriting or cleaning it; the cache may be evidence for how a prior graph was produced.

Verify outputs and evidence

A complete run should name the input root, exclusions, file types processed, local versus provider-backed stages, errors or skipped files, and the exact artifacts created:

• graphify-out/graph.html for interactive exploration.
• graphify-out/GRAPH_REPORT.md for prominent nodes and suggested questions.
• graphify-out/graph.json for persistent queries or MCP serving.
• graphify-out/cache/ for incremental file-hash state.

Open the HTML and report, sample important nodes against source, and confirm that the JSON retains provenance tags. Do not report the graph as complete when unsupported files, provider failures, or ambiguous relationships remain.

Troubleshoot safely

Know the current limits

• Language parsing, semantic extraction, transcription, clustering, and MCP features depend on separately installed package extras.
• Model-derived graph quality depends on provider availability and the selected model; graph proximity does not prove causation or ownership.
• The release evidence describes the shipped workflow but does not establish current upstream package behavior or end-to-end runtime parity.
• Stable and beta contain identical ak:graphify source.

Continue with ak:gkg for symbol navigation or ak:plan to turn verified graph evidence into a delivery roadmap.