Context Graph — Development skill for Claude Code

Context Graph · Knowledge Engineering Skill

This project provides a cross-runtime Knowledge Engineering Skill for Codex and Claude Code. It turns source material into traceable, reviewable knowledge and HTML-first projections that answer with evidence or deliberately abstain.

1. Overall structure


The system moves from a bounded question and preserved source material to canonical records, evidence checks, graph validation, projections, and measured retrieval. The final answer state is `answer`, `conflict`, or `abstain`.

2. Operation flow and order


The workflow has eight ordered stages:

1. Scope — define the question, domain, date, and acceptance criteria.
2. Collect — preserve documents, datasets, images, video, audio, or URLs.
3. Normalize — create stable Source, Evidence, Claim, Media, and provenance records.
4. Model — represent typed entities, relations, time, and ontology constraints.
5. Validate — check schemas, locators, hashes, freshness, reasoning, conflicts, and gold isolation.
6. Review — compare supporting and counter-evidence and decide accept, revise, hold, or reject.
7. Project — rebuild HTML, graph, search, reports, Context Packs, memory pointers, and optional Archify diagrams.
8. Measure — compare strategies on the same snapshot and question order.

The analysis/proposal role is read-only and returns an evidence-backed handoff. Together, analysis, proposal, review, and modification keep changes evidence-backed. The modification role applies only an approved request, preserves provenance, creates revisions, rebuilds projections, and verifies the result.

3. Installation and use

Codex or ChatGPT Skills

Download the latest release, then install the skill from:

context-graph/skills/knowle