Context Pack

Context Pack: A small, project-specific knowledge base for humans and agents.

A single, portable Claude Code skill — context-pack-setup — that provisions a complete knowledge base in any repository: it scaffolds a kb/ Open Knowledge Format (OKF) bundle, installs five primitive context-pack-* skills, wires a uv converter group, and interviews you for the KB's domain.

It is a Karpathy-style "LLM wiki": you curate sources, the LLM does the bookkeeping (summaries, interlinked pages, cross-references, health checks). Querying is plain markdown + grep + an index — deliberately no vector DB / RAG.

A knowledge base of 50 pages you trust beats 5,000 you have never read.

Install

./install.sh             # user scope:    ~/.claude/skills/context-pack-setup  (/context-pack-setup works in every repo)
./install.sh --project   # project scope: ./.claude/skills/context-pack-setup  (this repo only)

Or copy it by hand:

cp -r context-pack-setup ~/.claude/skills/

Re-running the installer overwrites the previous context-pack-setup install (it is tooling — safe to refresh).

Usage

In the repo where you want a KB, invoke the installer skill, then drive the workflow:

/context-pack-setup     # scaffold kb/, install the context-pack-* skills, wire converters, set the domain
/context-pack-add  <url-or-file-or-doi>   # capture a source → raw/ + a sources/ stub  (no synthesis)
/context-pack-ingest <slug>              # compile raw → literature note → wiki pages + links + index + log
/context-pack-ask  "what does the KB say about …?"   # cited answer, grounded only in the KB

Plus /context-pack-digest ("what's new since <date>") and /context-pack-lint (health-check the bundle).

Converters need their deps once: uv sync --group kb.

What /context-pack-setup creates

In the target repo:

kb/
├── README.md          # the contract (personalized with your domain); every context-pack-* skill reads it first
├── index.md  log.md   # reserved OKF files (no frontmatter): the routing index + activity log
├── inbox/             # staging for untriaged drops
├── raw/<slug>/        # immutable, faithful source payloads (content.md + meta.yaml)
├── sources/           # one literature note per source ("what does THIS source say?")
└── wiki/              # synthesized, cross-source concepts ("what do we know about X?")
    ├── concepts/  patterns/  tools/        # default categories (customizable at setup)
.claude/skills/
├── context-pack-add/  context-pack-ingest/  context-pack-ask/  context-pack-digest/  context-pack-lint/   # the five primitives
pyproject.toml         # gains a [dependency-groups] kb = ["docling>=2", "trafilatura>=1.12"]

/context-pack-setup is idempotent: it never overwrites existing files (cp -n), so re-running repairs a partial setup and leaves a complete one untouched. To re-personalize the domain, edit or delete kb/README.md and re-run.

Portability

The core is tool-agnostic: the contract (kb/README.md), the five primitive instruction files, the converter commands (uv run --group kb …, grep, shasum, cp), and the kb/ scaffold are plain markdown + POSIX shell with no Claude-Code-specific runtime variables. Any capable agent can operate the KB when pointed at these files.

There are two layers in a provisioned repo:

• kb/ is the portable knowledge artifact — any tool can read and operate on it.
• .claude/skills/ is the Claude Code adapter for driving it.

Only the context-pack-setup/SKILL.md installer itself is Claude-Code-specific (it uses ${CLAUDE_SKILL_DIR} to find its bundled assets and is invoked via slash command). Nothing it writes into your repo depends on a Claude-Code variable — so a user on another agent (e.g. Codex) gets the identical kb/ and can drive it via their own conventions.

Relation to dev-stack

This packages the knowledge-base machinery originally built into the dev-stack repo as a single standalone, portable skill so any repo can stand up the same setup with one command. The five primitive context-pack-* skills and the contract are byte-for-byte the dev-stack originals (the contract's domain is genericized into an interview prompt). dev-stack is the upstream source; context-pack is the distributable. Out of scope here (lives only in dev-stack's roadmap): autonomous bulk-fill (kb-bootstrap), parallel ingest, LLM-transcription fallback, and status-ladder enforcement.