Skip to main content
skill CX Codex

okf — Claude Code skill: what it does, how to install

Author, maintain, and consume Open Knowledge Format (OKF) knowledge bundles as markdown and YAML files directly inside your repository.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add scaccogatto/okf-skills --skill okf -a claude-code
Open on GitHub

What okf does

The skill guides an agent through authoring, updating, and reading knowledge bundles based on the Open Knowledge Format v0.2 specification. OKF stores project knowledge directly inside a repository, usually in the .okf/ folder. It uses standard Markdown files paired with YAML frontmatter. External databases are not needed. One file represents one concept.

The skill operates in three explicit modes: produce, maintain, and consume. The agent inspects and modifies files using tools like Read, Edit, Grep, and Bash. Initializing a clean bundle relies on uv run executing scripts/okf_init.py. It creates the root index.md and log.md.

Conformance hinges on one rule. Every non-reserved Markdown file must contain parseable YAML frontmatter with a non-empty type field. Other metadata fields remain optional. Without uv installed on your machine, the scaffolding script cannot run.

What it helps with

  • capture architecture decisions and API schemas in OKF bundles
  • scaffold a new .okf/ knowledge directory with root files
  • sync frontmatter metadata and log.md entries after code edits
  • link project documents using bundle-relative markdown references

Example prompts for the agent

  • Document this authentication service in OKF format under .okf/
  • Update our knowledge bundle to reflect recent codebase changes
  • Run okf in produce mode to capture this database schema

How to install okf

1Skills CLI (Vercel)

npx skills add scaccogatto/okf-skills --skill okf -a claude-code

2Manually: put the files in place

~/.claude/skills/okf/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/okf/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about okf

How to install okf in Claude Code?

Run npx skills add scaccogatto/okf-skills --skill okf -a claude-code in your terminal. This registers the okf skill directly for Claude Code.

What is an Open Knowledge Format bundle used for?

An OKF bundle organizes project knowledge into Markdown files with YAML frontmatter directly in the repository. It requires no external database and remains easy to parse for both human developers and AI agents.

What is the mandatory requirement for OKF markdown files?

Every non-reserved Markdown file must contain valid YAML frontmatter with a non-empty type property. According to the v0.2 specification, all other frontmatter fields are optional.

Source text

SKILL.md · MIT
---
name: okf
description: >-
  Author, maintain, and consume Open Knowledge Format (OKF) knowledge bundles —
  portable markdown + YAML frontmatter that both humans and agents read. Use when
  capturing project knowledge (services, APIs, schemas, metrics, runbooks,
  decisions) into an OKF bundle, when updating one after code or docs change, or
  when a repository contains an `.okf/` (or other OKF) bundle that should inform
  the task. Triggers on: "document this in OKF", "update the knowledge bundle",
  "capture this as a concept", or any work in a repo that has an OKF bundle.
user-invocable: true
argument-hint: "[produce|maintain|consume] [path]"
allowed-tools: Read Write Edit Grep Glob Bash
---

# Open Knowledge Format (OKF) skill

OKF represents knowledge as a directory of markdown files with YAML frontmatter.
It is minimal by design: no schema registry, no runtime, no SDK. Your job is to
produce, maintain, and consume OKF bundles **conformant with the spec**, not your
memory of it.

**Always read the canonical spec before non-trivial work:**
[reference/SPEC.md](reference/SPEC.md). It is the verbatim OKF v0.2 specification
and the source of truth for every rule below.

## The one hard rule

A bundle is conformant (§11) iff: every non-reserved `.md` file has a parseable
YAML frontmatter block, and every such block has a **non-empty `type`** field.
Everything else is soft guidance. Consumers MUST tolerate missing optional
fields, unknown types, and broken links — never reject a bundle over them.

## Conventions to apply

- **One concept = one file.** The file path (minus `.md`) is the concept ID.
- **Frontmatter:** `type` is required. Add `title`, `description`, `tags` when
  they aid consumption; add `resource` (a canonical URI) only for concepts bound
  to a real asset — omit it for abstract concepts.
- **Body:** prefer structural markdown (headings, tables, lists, fenced code).
  Conventional headings: `# Schema`, `# Examples`, `# Computation`.
- **Cross-links:** standard markdown links; prefer absolute bundle-relative
  form (`/services/auth-api.md`). A link asserts a relationship; its *kind* lives
  in the surrounding prose, not the link.
- **Reserved files:** `index.md` (directory listing, no frontmatter, except the
  bundle-root index, which may carry `okf_version` and this plugin's
  `upkeep: enforced` opt-in flag, plus `recursive: true` (or a depth `n`: levels below the root, `true` is 6) to make the hook
  watch nested git repos) and `log.md` (ISO-dated history, newest
  first). Never use these names for concepts.
- **What goes in `log.md`:** lifecycle events only, written in the bundle-root
  `log.md`: a bundle **created**, a concept **deprecated** or retired, one
  concept **superseding** another, concepts **regenerated** after their source
  changed, a **verification** pass. Routine edits to a concept do not get an
  entry: its `generated` and git already record them. A log is append-only, so
  never rewrite or drop an existing entry.

## The v0.2 families (all optional, all worth filling)

- **Trust (§5.2):** `generated: { by, at }` — who produced the current content
  and when. `verified: [{ by, at }]` — who confirmed it since (a bare mapping is
  one entry). Write `by` in the **actor convention** (§7): `<producer>/<version>`
  for an agent, `human:<id>` for a person, `process:<id>` for an automated job.
  Use `human:` whenever a person authored or signed off — consumers key trust
  tiers off that prefix.
- **Lifecycle (§5.4–5.5):** `status: draft|stable|deprecated` (absent means
  stable) and `stale_after`, an absolute ISO 8601 datetime (a bare date is
  tolerated), not a TTL.
- **Provenance (§5.1):** `sources: [{ id, resource, title, author,
  usage_count, last_modified }]` plus a `usage_window: { from, to }` sibling of
  `sources` framing every `usage_count` (an entry may carry its own to override
  it); a `usage_count` without a window warns. `resource` is required per entry
  and may be a URL, a bundle path, or a scope descriptor. Attribute a specific
  claim with a markdown footnote whose label is the source's `id`:
  `…sharded daily.[^ga4-schema]` plus a `[^ga4-schema]: …` definition. The label
  is the join key, it must match a `sources[].id`.
- **Attestation (§10):** a sanctioned computation is its own concept,
  `type: Attested Computation`, carrying `runtime` (required), `parameters`,
  `executor`, `attester`, and the computation itself under `# Computation` (or a
  `computation:` path). Concepts that need the value link to it. Never inline a
  number's SQL into the concept that narrates it.

**Reading a v0.1 bundle?** Two constructs were superseded (§13.1): `timestamp`
is now `generated.at`, and a body `# Citations` list is now `sources`. Read both,
write v0.2 — and when you touch a legacy concept in **maintain** mode, migrate
its frontmatter as part of the edit. The validator warns on both.

Templates to copy: [concept](templates/concept.md), [index](templates/index.md),
[log](templates/log.md).

## Default bundle location

Use `.okf/` at the repository root unless the project already uses another
location. Commit it alongside the code it describes — knowledge as code.

## Modes

### produce — create or extend a bundle

**Starting a brand-new bundle?** Use the init fast-path instead of hand-writing
the first files — it scaffolds a conformant `index.md`, `log.md`, and a
`getting-started.md` concept with full recommended frontmatter in one shot:

```bash
uv run "${CLAUDE_SKILL_DIR}/scripts/okf_init.py" <target-dir> [--title "..."]
```

It refuses to touch a directory that already has `.md` files unless `--force`
is given. Then extend it:

1. Read [reference/SPEC.md](reference/SPEC.md).
2. Pick the source(s): **code** (derive concepts from source, READMEs,
   docstrings, config), **docs/wiki** (distill pages into concepts, record the
   originals in `sources`), **manual** (decisions, playbooks, metrics).
3. Choose a directory layout by domain (e.g. `services/`, `datasets/`,
   `decisions/`). One concept per file.
4. Write each concept from [templates/concept.md](templates/concept.md): set a
   descriptive `type`, fill recommended fields, record `generated` and the
   `sources` you actually read, cross-link related concepts.
5. Add/refresh `index.md` per directory (and `okf_version: "0.2"` in the root
   index). A new bundle gets a dated **Creation** entry in `log.md`. Extending
   an existing one does not.
6. Validate (see below). Fix every error before finishing.

### maintain — keep a bundle in sync with reality
1. Identify which concepts the change affects (search by `resource`, path, or
   topic). This bookkeeping is exactly what agents are good at — touch every
   affected file in one pass.
2. Update the body and `generated.at` (with your own actor in `generated.by`);
   fix or add cross-links; create new concepts for new assets; mark removed
   assets `status: deprecated` and note the deprecation in `log.md` rather than
   silently deleting context. Facing a whole v0.1 bundle rather than a stray
   field? Do not hand-edit it — run the validator's `--migrate` once.
3. Update the relevant `index.md` files. Append a dated `log.md` entry only if
   the change is a lifecycle event (see *What goes in `log.md`*).
4. Validate.

### consume — use a bundle as context
1. Read the bundle-root `index.md` first for progressive disclosure, then follow
   links only into the concepts relevant to the task.
2. Weigh what you read: `status: draft`/`deprecated`, a `stale_after` already
   past, or no `verified` entry all mean "check before relying on this". Treat
   broken links as not-yet-written knowledge, not errors.
3. Need a number an `Attested Computation` covers? Run *its* computation with
   values bound to the declared `parameters` — never write your own query.
4. If you learn something durable while working, switch to **maintain** and
   write it back.

## Validation (do this before declaring done)

Never eyeball conformance — run the deterministic checker. Invoke the companion
**`validate`** skill (`/okf:validate <bundle-dir> --strict`), which ships the
checker. If that skill is not installed, run it directly:

```bash
uv run "${CLAUDE_SKILL_DIR}/../validate/scripts/okf_validate.py" <bundle-dir> --strict
```

Resolve every `ERROR` (hard §11 failures). Under `--strict` warnings also fail
the run (exit 1), so clear them too; drop `--strict` (or use `--max-warnings N`)
if you only want §11 conformance gated.

© scaccogatto · MIT · scaccogatto/okf-skills

Similar skills

Everything for Writing and docs →

From the same set2

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor