mcpbeat

Okf Wiki

jamditis/okf-wiki

>- Scaffold a new Open Knowledge Format (OKF) knowledge base and populate it frontmatter, a spec, a validator, and session-start hooks that orient Claude on the knowledge base before it works. Use when the user wants to start an OKF atlas/wiki/knowledge base, build one from existing docs, plans, notes, or a repo, structure docs as one-concept-per-file with provenance, or initialize OKF in a repo (optionally into its GitHub wiki).

109k tokens
context cost
the whole folder, loaded on every use
114
files
ships runnable scripts
0
copies elsewhere
how many repositories repackaged it
351
stars on the repo
on the repository, not the skill itself

Install

one command, takes just this skill from the repository
npx skills add https://github.com/jamditis/claude-skills-journalism --skill okf-wiki

The instruction itself

13 sections, as written by the author

okf-wiki: scaffold an Open Knowledge Format knowledge base

OKF (Open Knowledge Format) stores knowledge as small markdown files: one concept per

file, each carrying its own provenance in YAML frontmatter, with directory index.md

files for navigation and a validator that enforces the contract. It is built for

knowledge bases that both people and agents read and edit — newsroom institutional

memory, a research atlas, a team's decision log, an infrastructure map.

This skill scaffolds a conforming OKF project and validates it. The format contract is

in spec/SPEC.md (in this skill's directory) — read it before changing structure.

When to use

  • The user wants to start an OKF knowledge base, atlas, or wiki.
  • They want docs structured as one-concept-per-file with provenance, not prose pages.
  • They want to "initialize OKF" in a repo, optionally publishing into its GitHub wiki.

Start here: scope the wiki with the user

Before you scaffold anything, settle four things with the user. They shape what gets created and

how it is published, and they are awkward to retrofit once concepts exist. Ask with AskUserQuestion

rather than in prose, in two steps: the first three questions in one call, then the publish question

as a follow-up call only if the audience came back public or both (it does not apply to an

internal-only wiki, and its relevant options depend on that answer, so it cannot share the first

batch). Infer the title from the repo or project and confirm it. Skip any question the user already

answered in their request — do not re-ask what they have told you.

  • Audience — who reads this wiki? This answer sets the others:
  • Internal (agents and teammates): the orientation hooks earn their keep, so keep them on.

The bundle may hold infrastructure detail, so it usually lives in a private repo. The in-repo

bundle/ is the source of truth.

  • Public (people browsing): readability and secret-scrubbing come first; the hooks matter

less, since people read it and agents do not. Plan a published view (see Publish below).

  • Both: the in-repo bundle/ is the source of truth with hooks on for agents, plus a

published view for people. Default here when the user is unsure.

  • Title and sections — the knowledge-base title (infer it, then confirm) and the starting

sections. Offer sections as a use-case preset, not a blank prompt:

  • Newsroom institutional memory: people, orgs, sources, decisions, beats
  • Research atlas: concepts, sources, methods, findings
  • Infrastructure or fleet map: machines, services, networks, credentials, processes
  • Decision log: decisions, context, events

The chosen title and list feed --title and --sections below; the user can edit the list.

  • Populate now or later — author concepts now from existing material (a repo, docs, notes, or a

URL: gather it and enter the authoring loop after scaffolding), or scaffold an empty tree the user

fills in later.

  • Publish target — a follow-up AskUserQuestion call, made only after the audience comes back

public or both (skip it entirely for an internal-only wiki):

  • In-repo bundle only (default): the validator and relative links work directly, with no

extra surface to maintain. Right for most wikis.

  • GitHub wiki: an optional reading surface. Advanced and manual — see "Optional: publish into

a GitHub wiki" below, bootstrapped with scripts/gh-wiki-bootstrap.py.

  • GitHub Pages: a browsable site rendered from the bundle. Not built yet — treat it as a

goal and keep the in-repo bundle as the source of truth.

Carry the answers into the scaffold command (the title and sections, plus --no-hooks if the user

opts out of the hooks for a public-only wiki) and into the populate step. The audience answer is

also the visibility decision the "Before finishing" section asks you to make deliberately — you are

making it here, up front, where it can steer the rest of the setup.

What gets created

scripts/scaffold.py writes a project that passes its own validator by construction:

<target>/
  SPEC.md                 the OKF format contract
  README.md               how to use and validate the bundle
  scripts/validate.py     the validator
  .claude/                Claude Code adapter: session-orientation hooks
    settings.json         registers the hooks (Claude Code approves them once)
    hooks/okf-anchor.py   SessionStart: load the index into context
    hooks/okf-orient.py   PreToolUse: gate the first action on orientation
  bundle/                 the OKF bundle (the validated tree)
    index.md              carries okf_version: "0.3" by default; "0.4" with --trust-signals
    <section>/
      index.md
      example-concept.md  a starter concept with full frontmatter

Docs and tooling sit at the project root; only bundle/ is validated. Keep them

separate — the validator treats every non-reserved .md inside the bundle as a

concept that needs frontmatter, so a stray SPEC.md inside bundle/ would fail.

The .claude/ hooks sit outside bundle/, so they never trip the concept checks.

How to run it

${CLAUDE_SKILL_DIR} below is this skill's own directory (the folder holding this

SKILL.md). Claude Code substitutes it with the real absolute path before you run the

command, so it works regardless of the current directory. On Windows, use python instead

of python3 (stock Windows has no python3). The --title and --sections come from the

onboarding answers above, and --no-hooks only if the user opted out. Scaffold into a new

directory; it validates automatically at the end:

python3 "${CLAUDE_SKILL_DIR}/scripts/scaffold.py" ./my-knowledge-base \
  --title "Team knowledge base" \
  --sections concepts,services,decisions

Default section is concepts. Use --force to write into a non-empty directory,

--no-validate to skip the validation run, and --date YYYY-MM-DD to set the sample

frontmatter date. The session hooks are written by default; --no-hooks skips them and

--hooks-os posix|windows overrides the auto-detected launch command (see below).

Validate any time, from the scaffolded project root (use python on Windows):

python3 scripts/validate.py --bundle bundle    # must exit 0

Populate the bundle: author concepts from existing material

Scaffolding leaves an empty tree with one placeholder concept. The usual next request

— "here are my docs / plans / notes / repo, build the wiki" — has no importer script, and

can't have one: deciding what counts as a single concept, writing its one-line description,

choosing its type, and pointing source at real provenance is judgment work, not a

mechanical transform. So you (Claude) author the concepts directly, in this loop:

  • Gather the source. Read what the user pointed you at — a file, a folder, a repo, or a

URL (fetch a URL first). Skim the whole thing before writing anything, so you can see the

natural concept boundaries.

  • Decide concept boundaries. One file is one concept: one thing a reader would look up on

its own (a service, a decision, a path, a person, an event). Split a doc that covers five

things into five concepts; merge fragments that only mean something together into one. A

heading is a hint, not a rule — do not blindly map one ## to one file.

  • Draft each concept at bundle/<section>/<slug>.md with the full frontmatter. Read the

bundle-root index.md before writing so the verification key matches its declared format:

use verified for okf_version 0.1 through 0.3; use verified_on for okf_version 0.4.

Emit that exact key with type, title, description, source, timestamp, tags:

  • type from the vocab. Infrastructure: Machine, Network, Service, Session, Project,

Repo, Credential, Path, Process. Domain-neutral: Concept, Decision, Event, Person,

Org, Source. Plus Reference (the catch-all). The set is closed; an unlisted type fails.

  • description is one line. source — quote every element — points at where the fact

actually came from (the origin file path, URL, command, or event), not at this skill.

  • Set timestamp to today. verified/verified_on is the date the fact was last confirmed

true — set it by how you came to know it, not reflexively to today:

  • You re-checked it against reality now, or the user is the authority for it (a decision,

preference, or intent they state in this session): today.

  • The user is recalling external or system state (a spec, a path, a config): their memory is a

source claim, not a re-check, so date it to when that state was last checked or to the

recollection's own date — not today just because it came up now.

  • It was copied from a dated source without re-checking: the date it was last known true (the

source's own date), not today.

  • It came from an undated record you cannot re-confirm (a memory file, an old conversation):

the oldest date you can evidence — file timestamp, introducing commit, or the date it was

said — never today. If you cannot evidence any date at all, it is not yet a verifiable fact;

find a datable source or leave the concept out.

When the date is uncertain, round it down: an older verified correctly reads as "may be

stale, re-check," while today reads as "just confirmed." The frontmatter date is the contract;

a caveat in the body does not undo an overstated value, because the validator and tools read

only the date.

  • Strip secret values as you go: a credential concept names the key and its retrieval path,

never the value. The validator fails the build on a leaked secret.

  • Place and link. Put each concept in the right section (create sections as needed), add a

bullet for it to that section's index.md, and cross-link related concepts with relative

text links — not [[slug]] wikilinks. [[slug]] is the auto-memory idiom; the

OKF validator rejects it and never resolves it, so a typo'd or deleted reference passes silently.

When you create a new section, also link it from the bundle-root index.md — that root is the

navigation map the session anchor loads, so a section missing from it is invisible to orientation

even though validation still passes.

  • Clear the placeholder. If you scaffolded fresh, delete the starter example-concept.md (and

its bullet in the section index.md) once real concepts exist — otherwise the sample ships in

the finished wiki and still passes validation.

  • Validate in a loop. Run python3 scripts/validate.py --bundle bundle, fix what it

reports, repeat until it exits 0. Unquoted source elements and missing frontmatter keys are

the common failures. Author in batches and validate between them rather than writing fifty

files and debugging the lot.

When the source is already OKF

If the user points you at an existing OKF bundle (e.g. an upstream example: an index.md

carrying okf_version plus concept files with frontmatter), you are adopting it, not importing

it. Copy or clone the tree in, point the validator at the new root, and fix any links that broke

in the move. To keep it as its own area beside other content, give it a uniquely named top

directory, then create one combined-root index.md that carries okf_version and strip the

frontmatter from each adopted bundle's own root index.md, turning it into a normal section index

(the validator allows okf_version on the one combined root only; a nested index.md that still

carries it fails validation). Write cross-links as relative paths and validate the combined root.

Re-authoring an already-conforming bundle into your own concepts is wasted work; only reshape it if

that is the actual goal.

The format, briefly

Full contract in spec/SPEC.md. This spec is a strict fork of Google's upstream OKF: it

requires all seven frontmatter keys, uses a source list in place of upstream's resource

and # Citations, adds a verification-date key, closes the type vocab, and enforces link

resolution. spec/SPEC.md ("Relationship to upstream OKF") lists every difference. The

load-bearing rules:

  • Required frontmatter on every concept: type, title, description, source, the

version-specific verification key described above, timestamp, tags. type is one of:

Machine, Network, Service, Session, Project, Repo, Credential, Path, Process

(infrastructure); Concept, Decision, Event, Person, Org, Source (domain-neutral); or

Reference (catch-all).

  • Quote every source element — source pointers carry # and : which break YAML

if unquoted. source: ["README.md", "issue #445"].

  • verified/verified_on is the date the fact was last confirmed true — a re-check

against reality, or the user stating a fact they are the authority for (a decision, a

preference); a fact they merely recall about external state is a source claim, not a

re-check. timestamp is when the concept was authored/updated. The verification date is

ISO YYYY-MM-DD; timestamp may also be a full ISO 8601 datetime in 0.3 and 0.4. See

the authoring loop above for the full date rules.

  • No secret values, ever. A credential concept documents the key name and retrieval

path, never the value. The validator fails the build on a leaked secret.

  • index.md and log.md are reserved — no frontmatter (except the bundle-root

index.md, which carries okf_version only).

Optional: upstream v0.2 trust/provenance signals

Upstream Google OKF v0.2 (July 2026) added an optional vocabulary for a consumer to judge a

concept before reading it: generated (who/what produced it), verified (a list of

independent confirmations, not this fork's own single-date field), sources (structured,

per-pointer credibility signals), status (draft/stable/deprecated), stale_after (an

absolute expiry date), and an Attested Computation type for a sanctioned, checkable

computation. None of it is required, and a bundle that adopts none of it is unaffected.

Scaffold a project with these enabled — scaffold.py <target> --trust-signals — and the

bundle declares okf_version: "0.4", with verified renamed to verified_on in the

required set (freeing verified for the new shape; see spec/SPEC.md's "Trust and

provenance" section for the full field contract and the reasoning behind the rename).

Attested Computation is likewise a 0.4-only type. Without the flag, scaffolding is

unchanged from before this vocabulary existed.

Session hooks

A scaffolded project ships a .claude/ with two hooks so any Claude session opened in it

starts from the bundle, not from memory:

  • okf-anchor.py (SessionStart) prints the bundle's root index into the session context.
  • okf-orient.py (PreToolUse, no matcher) blocks the first action of the session once,

until Claude confirms it read the index, then unblocks for the rest of the session. It is

inert outside an OKF bundle and fails open on any error, so it never wedges a session.

Both are one cross-platform python3 script. The scripts are identical on every OS; only the

interpreter in .claude/settings.json changes: python3 on macOS/Linux, python on

Windows. scaffold.py auto-detects the OS; --hooks-os posix|windows forces it.

Claude Code treats a checked-in .claude/settings.json as untrusted, so the first time the

project is opened it asks the user to approve the hooks; they run automatically after that.

To turn them off, scaffold with --no-hooks, or delete .claude/ (or set disableAllHooks)

in an existing project.

Client boundary

The portable OKF surface is SPEC.md, requirements.txt, scripts/validate.py, and the

bundle/ tree. The generated README.md documents both that shared surface and any enabled

client adapter. The three generated .claude/ files are a Claude Code adapter, not part of

the OKF format and not shared Codex behavior. Codex does not read them as project configuration,

and this skill must not claim that their SessionStart or PreToolUse lifecycle runs there.

The general onboarding route above still names Claude Code's AskUserQuestion and

${CLAUDE_SKILL_DIR} surfaces. The recorded Codex pilot pre-set every onboarding choice and

used an explicit project-relative installed path; it does not establish that the unadapted

general route is portable.

For a mixed Claude Code and Codex project, keep .claude/ so Claude Code can request trust

and use the hooks; Codex leaves it inert. For a Codex-only project, pass --no-hooks while

scaffolding or delete .claude/ afterward. Either choice leaves the portable bundle and

validator unchanged.

Optional: publish into a GitHub wiki

OKF lives best as in-repo files (the validator and relative links work directly). A repo's

GitHub wiki is an optional reading surface, and wiring it up is an advanced, manual step —

most users should skip it and keep the bundle in-repo.

A wiki with zero pages has no git repo to push to and no API, so the very first page must be

created through the web UI. scripts/gh-wiki-bootstrap.py automates that one step, but it

drives a real logged-in browser, so it needs two things you provide yourself (a GitHub PAT

does not work — wiki pages are a web-UI-only surface):

  • Playwright with Chromium installed: pip install playwright && playwright install chromium.
  • A saved GitHub web session: a Playwright storageState JSON, captured from a browser

where you have already logged into GitHub. The script reuses that session; it does not log

in for you. Pass its path with --state (default: ~/.cache/gh_state.json).

python3 "${CLAUDE_SKILL_DIR}/scripts/gh-wiki-bootstrap.py" owner/repo --state path/to/gh_state.json
# then: git clone https://github.com/owner/repo.wiki.git and push your pages

Note the impedance: GitHub wikis are flatter than an OKF tree and use [[WikiLinks]], so

OKF's nested directories and relative links need adapting for the wiki surface. Treat the

wiki as a published view, not the source of truth. (v0.1 ships the bootstrap step; an

automatic bundle-to-wiki sync is not built yet.)

Before finishing

  • Run the validator and confirm it exits 0.
  • Confirm the visibility you set during onboarding still fits what got authored: a bundle that

ended up documenting real infrastructure is usually internal. OKF takes no position; you must.

How to use it

Copy the folder

Take jamditis/okf-wiki from the repository into ~/.claude/skills for personal use, or into .claude/skills inside a project.

Check the name does not clash

The agent identifies a skill by the name field in its header. Two skills with the same name cannot sit side by side — one of them will be ignored.

Install what it needs

The instructions reference pip. Without those the skill loads but fails at the first command.