Ingest — Assessed openwiki (langchain-ai), a new LangChain CLI that auto-documents codebases for agents. Verdict: adjacent lane, not a format competitor — plain markdown (no frontmatter/type), codebase-docs only, no trust model, regenerates from source. Matters for two reasons: (1) LangChain entering hardens the "agent-maintained wiki" category and our positioning as the OKF/general-knowledge/trust-modeled option; (2) its agent prompt is a proven source of kb-ingest techniques — git-evidence grounding, plan-then-write, read-only research subagents — now recorded in Skill Design. Added ecosystem landscape category 7 (codebase-doc generators).
Update — Pre-publish pass. Generated viz.html for this bundle (committed at bundle root, like the OKF samples). Anonymized the two implementation prior-art pages for public release — a personal work wiki and an OKF-native agent — removing internal project/program names, private repo URLs, and the author's home path while keeping the structural lessons. Added repo-level README.md (mattpocock-style: why → install → skills-by-invocation → self-documenting bundle), LICENSE (MIT + Apache-2.0 attribution for the vendored SPEC), and .gitignore. Prepared to publish as stjbrown/agent-knowledge.
Creation — Built kb-visualize (user-invoked) — completes the family (6/6). A bundled scripts/graph.py (no deps) extracts the deterministic graph model — nodes by type, edges + cited_by backlinks from cross-links, frontmatter + body per node; the agent then authors the view (adaptive, can scope to a subgraph), rather than running a fixed template. Key branch: render as native UI where the host supports it (Claude/Codex Desktop), else a self-contained HTML file (Cytoscape + marked via CDN, model inlined). View mirrors a conformant OKF viewer: colored force-graph, detail panel with in-view link navigation, backlinks, search, type filter. Read-only (no log entry from the skill itself). Tested graph.py on this bundle (46 nodes/325 edges/5 types; hub = operations/ingest at 25 inbound) and the example bundle. Family fully cross-links; manifest lists all six.
Creation — Built kb-lint (user-invoked) — the maintenance health-check. Two passes: a deterministic conformance check via a bundled scripts/conformance.py (self-contained python3, zero-install — reuses the mechanical §9 check without depending on an external validator; ERROR = hard §9 failure, warn = broken links/non-ISO dates which never fail per §5.3) and a drift audit (contradictions, stale claims, orphans, missing cross-refs, coverage/provenance gaps) with an exhaustive "every check, whole bundle" completion criterion. Severity-grouped report + append-only log summary; fix mode repairs only safe/unambiguous issues and flags meaning-level ones for ingest (never edits a claim in place). Tested the script against this bundle (55 files/46 concepts, conformant) and the example bundle — both clean. Registered in the manifest. Family now 5/6; only kb-visualize remains.
Update — Ran a writing-great-skills audit pass over the built skills. Fixed one real duplication smell: kb-ingest step 5 restated the trust model's supersede/conflict/event rules that already live in kb/reference/trust-model.md (the single source of truth) — collapsed it to execute the plan, following the trust model for mechanics. Confirmed kb-query step 3's read-time rules are not duplication (the trust model covers only the writing side; read-time behavior has no other home). No no-ops/sprawl found elsewhere worth cutting.
Creation — Built kb-ingest (model-invoked) — the core differentiator no competitor has: read a raw source once → extract → integrate across the bundle under the trust model. 9 steps, each with a checkable completion criterion (the premature-completion defense for this long sequence): locate bundle + read its schema layer → read/classify source → plan (openwiki's plan-then-write; the plan becomes the checklist later steps complete against) → store source as a Reference (provenance, N:1) → integrate under the trust model (supersede/conflict/events-additive, both-direction cross-links) → re-synthesize overviews → update indexes → append log → retire source + commit. Read-only research subagents allowed for large batches; only the main run writes. Registered in the manifest. All kb-ingest cross-links resolve; remaining forward-refs are kb-lint/kb-visualize.
Update — Fixed the kb hub's ambient-consult line to route to kb-query rather than describe the read inline (single-source-of-truth: reading logic lives only in kb-query). Built kb-query (model-invoked): two modes (explicit question + ambient consult for another task), progressive-disclosure navigation, currency/conflict-aware reading (follow superseded_by; answer contested questions with nuance), cite specific concepts, and file valuable answers back so queries compound — each step with a checkable completion criterion. Registered it in .claude-plugin/plugin.json. Remaining skill cross-links are forward-refs to kb-ingest/kb-lint/kb-visualize.
Creation — Built the skill foundation under skills/: the kb hub (model-invoked router — SKILL.md, reference/SPEC.md vendored verbatim, reference/glossary.md, reference/trust-model.md, templates/{concept,index,log}.md, and a self-contained conformant example-bundle/ that doubles as kb-init's seed) and kb-init (user-invoked scaffolder: default knowledge/, custom path, multi-bundle aware, writes the schema layer as spec/ concepts, print-a-prompt domain tailoring). Added .claude-plugin/plugin.json (skills.sh + Claude Code plugin, mattpocock/skills layout). Example-bundle verified conformant; remaining cross-links are forward-refs to the not-yet-built action skills.
Update — Settled the three open build decisions in Skill Design: (1) distribution → skills.sh single-plugin layout (mattpocock/skills model — skills/<name>/SKILL.md + .claude-plugin/plugin.json, no build step; family ships as one plugin so action skills reach the kb hub's shared reference); (2) schema layer → in-bundle spec/ concepts written by kb-init; (3) trust model → the full opinionated OKF-native-agent trust model. Build starts: kb hub → kb-init → kb-ingest → kb-query → kb-lint → kb-visualize.
Creation — Filed Skill Design — the kb-* family: the build plan, designed with the writing-great-skills framework. Decided kb- naming ("knowledge bundle," OKF's own term; no user-facing "okf"), a kb hub skill (router + single source of truth holding the vendored spec/glossary/trust-model/templates/example bundle), and the invocation split — model-invoke only the daily verbs (kb-ingest, kb-query) + the hub; user-invoke the deliberate ones (kb-init, kb-lint, kb-visualize, kb-search) → three descriptions in context, not seven. Records leading words, the ingest premature-completion guard, build-vs-reuse, build order, and the three open decisions (distribution, schema-layer location, trust-model strictness).
Update — Audited references/ for source-vs-synthesis (per §8 + #91). Found two pages that were our analysis, not source material, mis-typed as Reference: moved them to a new design/ section and retyped to Concept — references/okf_spec_evolution.md → design/spec_evolution.md, references/okf_sample_bundles.md → design/sample_bundle_lessons.md (dropped their resource:, since a synthesis concept isn't bound to the asset it cites). references/ is now pure source pointers; design/ holds our format/build analysis; ecosystem/ holds analysis of other projects. Fixed all inbound/outbound links and added design/index.md.
Update — Read the full SPEC.md end-to-end (previously worked from a fetch-summary) and reviewed discussion #91 on references. Corrected a gap in citations: §8 explicitly allows citations to point into a references/ subdirectory that mirrors external material as first-class concepts, and #91's canonical-source-vs-derived-text discipline. Documented both there and in OKF Spec Evolution §6. Confirms the infographic's references/ placement is spec-backed, not a stylistic choice.
Ingest — Added a two-panel "OKF vs. RAG" infographic as a marketing/explainer asset. Stored both images under assets/images/ (immutable), created OKF vs. RAG infographic with descriptive alt text and honest unverified-source provenance, and cross-linked it from RAG vs. LLM Wiki. Source unverified (social-media capture) — confirm rights/attribution before public use.
Update — Converted the entire bundle from absolute (/…) to relative cross-links (473 links across 51 files) so it renders correctly on GitHub and for any non-OKF-aware reader, aligning with the reference agent and sample bundles (see OKF Spec Evolution §1). Rewrote cross-linking §5.1/§5.2 to present relative as the practical recommendation, and recorded the resolved/pending decisions (trust-model field names, README/LICENSE tolerance) in the spec-evolution page.
Creation — Bootstrapped the bundle. Created the root index and this log.
Creation — Wrote the concepts section: LLM Wiki, three-layer architecture, compounding artifact, progressive disclosure, RAG comparison, Memex, knowledge bundle, concept document.
Creation — Wrote the spec section: one concept per OKF v0.1 spec section (motivation, terminology, bundle structure, reserved filenames, frontmatter, body, cross-linking, index files, log files, citations, conformance, versioning).
Creation — Wrote the operations section (ingest, query, lint), the implementations section (a personal work wiki, an OKF-native agent), and the references section (OKF spec, OKF README, Karpathy gist, qmd).
Creation — Added the ecosystem section, synthesized from the ~900 comments on Karpathy's gist: a landscape survey of 200+ projects, a critiques page (truth maintenance, token cost, markdown-vs-database), and six featured projects (OmegaWiki, karpathy-llm-wiki, wiki-skills, okf-harness, synthadoc, commonplace).
Creation — Added OKF Sample Bundles reference, capturing design lessons from Google's ga4/stackoverflow/crypto_bitcoin bundles (minimal per-directory indexes, no log.md in generated snapshots, relative links, non-atomic concepts).
Creation — Added three OKF-native competitors/peers to the ecosystem: okf-skills (scaccogatto) — our closest competitor; openknowledge (openknowledge-sh) — Go CLI + registry + viewer; and kiso (oak-invest) — a Java bundle publisher. Updated the landscape's "where OKF sits" analysis: the native cohort is real and our differentiator must be the maintenance loop + portability seam, not another author/validate/visualize toolkit.
Creation — Filed Competitor Comparison: a feature-by-feature analysis of okf-skills vs. openknowledge vs. our plan, based on reading their actual SKILL.md files, templates, and validation code. Conclusion: reuse their deterministic validation + visualization/export; differentiate on raw-source ingest, the trust model, and the schema-layer portability seam. Proposed a three-skill scope (okf-init, okf-ingest, okf-lint).
Creation — Filed OKF Spec Evolution from reviewing the open PRs on knowledge-catalog. Key findings: link form is being reversed (relative recommended over absolute, PR #165 — flags an open decision for this bundle); only type is truly required (#145/#161, validates our checker); a trust/provenance axis is emerging (#58 §12 trust&safety, #159 reliability, #50 sources) that aligns with our differentiator and may standardize field names; hand-authored domain bundles with per-bundle spec/ files are being legitimized (#144 cricket), validating our schema-layer approach. Added an in-flux note to cross-linking.