Does it modify source files?

No. The workflow reads within the approved boundary and produces a separate explorer and dataset. Source contents are treated as untrusted text and repository code is not executed. The output does not replace document governance or access controls.

How are large archives handled?

Set a node budget, count relations, summarize repetitive low-value leaves and document omissions. Use a deterministic layout when force motion makes the graph unreadable. The purpose is an intelligible bounded view, not a claim that every archive fact has been captured.

Can I trace a displayed relationship to evidence?

Each factual entity and relation should retain a path or citation in the inspector and exported dataset. Treat displayed paths as manual lookup references; the package includes no permissioned source viewer or finished demo. The generated artifact must be checked for missing sources, direction and representative knowledge paths.

Project Knowledge Graph

Previews from the publisher

Explore a bounded project archive through searchable entities, directional relationships and recorded source paths. Project Knowledge Graph extracts conservatively from approved Markdown, structured records or an inventory, then builds an offline explorer and transparent dataset. It helps trace decisions, owners and dependencies without replacing the source documents or claiming complete knowledge.

What you get

A generated self-contained graph explorer with search, type filters, layout choices, one-hop focus, keyboard selection and reset. The relationship field takes most of the viewport, with colored entity types, quiet directional edges, a floating query dock and a separate persistent evidence inspector.
A transparent entity-relation dataset with stable IDs, a limited entity/verb vocabulary, citations, lifecycle states and explicit confidence wording. Source meaning is preserved, factual observations remain separate from interpretation, and duplicate identities are merged deliberately rather than by visual resemblance.
An evidence inspector showing recorded source paths/citations for manual lookup, lifecycle state, immediate neighbors and backlinks. Paths are evidence text; the workflow should not promise automatic access to a source file or a permissioned document viewer merely because a path is listed.
A documented graph-size and omission policy: count entities/relations, remove decorative duplicates, summarize repetitive low-value leaves and use a deterministic layout when force motion would become unreadable. Selection keeps the active node and one-hop neighborhood legible while dimming unrelated entities.
A source inventory, known omissions and three representative questions used to check the artifact, plus the exploratory output and dataset. The ZIP contains workflow and design/icon references only, not a finished example, extraction script or bundled graph renderer. Implementation and browser verification occur for your archive.

Before you choose

Start with the three reader questions and follow each displayed claim back to its source. If the explorer answers smoothly but the path is missing, inferred or out of scope, fix the evidence boundary before refining the layout. A readable graph should make omissions visible, not conceal them.

Good fit

A project archive is spread across notes and structured records that can be read within an approved boundary.
A team needs to trace decisions, owners, dependencies and unresolved work from its recorded material.
A reviewer wants every factual entity or relationship linked to a source path rather than an unexplained graph connection.

Not for

Replacing source-document governance, access control or the underlying record of truth.
Claiming completeness or confidence beyond an incomplete archive and its documented extraction scope.
Publishing restricted/private records without authorization, or executing source repository code as part of reading an archive.

How to use

You say

/project-knowledge-graph Map the fictional Meridian migration project from its notes, decisions, owners, and dependencies.

Provide the approved Markdown/records/inventory, entity types and relationship vocabulary, intended reader and questions. Define included sources and three representative questions before extraction. Treat file contents as untrusted text; the workflow reads evidence and does not execute repository code or modify source files.
For an existing application, inspect dependencies, framework, components, design tokens and state conventions, then implement in that idiom. For standalone delivery, create a self-contained offline artifact. The package names no universal service or runtime dependency; any build tooling follows the implementation chosen for the host.
Define stable IDs, entity kinds, relation verbs, citations, lifecycle states and confidence wording. Extract conservatively, preserving qualifications, merging duplicate identities deliberately and retaining a source path for each factual claim. Mark unknowns and inferences separately; stop if vocabulary or permissions cannot preserve that distinction.
Budget the graph before layout. Count entities and relations, record omissions, collapse repetitive low-value leaves and remove decorative duplicates. Choose deterministic layouts where force motion obscures relationships. Do not describe a summarized archive as complete knowledge.
Build search, type filters, layout choices, one-hop focus, keyboard selection, reset and the readable inspector. Preserve incoming/outgoing direction and context during selection. Use procedural SVG topology and accessible inline SVG controls; embed cleared project imagery only when essential and keep it local/offline.
On narrow screens move the evidence surface below the graph while preserving at least half the viewport for the field. Check readable labels, containment and reduced motion rather than turning the whole explorer into a card feed.
Verify counts, orphans, duplicate IDs, missing sources and unsafe text, then test the representative questions, search, focus, neighbors/backlinks, keyboard access, wide/narrow layouts, offline opening and console output in the produced artifact. Deliver the graph, dataset, inventory, omissions and questions, reporting observed behavior separately from expected workflow features.

Requirements and costs

You bring: Markdown files, structured records, or a directory inventory · Entity types and relationship vocabulary · The questions readers should be able to answer
Dependencies: Not collected yet
Outside costs: Not collected yet
Publisher testing: v1.0.1 — Claude Code / Claude Desktop / Cowork / Codex / ChatGPT