-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathviz.html
More file actions
2102 lines (1601 loc) · 178 KB
/
Copy pathviz.html
File metadata and controls
2102 lines (1601 loc) · 178 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<title>agent-knowledge — bundle graph</title>
<script src="https://cdnjs.cloudflare.com/ajax/libs/cytoscape/3.30.2/cytoscape.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
<style>
:root{--bg:#0d1117;--panel:#161b22;--border:#30363d;--fg:#e6edf3;--mut:#8b949e}
*{box-sizing:border-box}html,body{margin:0;height:100%;font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Helvetica,Arial,sans-serif;background:var(--bg);color:var(--fg)}
#app{display:flex;height:100vh}
#graph{flex:1;min-width:0}
#side{width:420px;border-left:1px solid var(--border);background:var(--panel);display:flex;flex-direction:column}
#controls{padding:12px;border-bottom:1px solid var(--border);display:flex;flex-wrap:wrap;gap:8px;align-items:center}
#controls input,#controls select{background:var(--bg);color:var(--fg);border:1px solid var(--border);border-radius:6px;padding:6px 8px;font-size:13px}
#search{flex:1;min-width:120px}
#detail{padding:16px;overflow:auto;flex:1}
#detail h1{font-size:20px;margin:.2em 0}
#detail .meta{color:var(--mut);font-size:13px;margin-bottom:12px}
#detail .tag{display:inline-block;background:#21262d;border:1px solid var(--border);border-radius:10px;padding:1px 8px;margin:2px 4px 2px 0;font-size:12px}
#detail .body{border-top:1px solid var(--border);margin-top:12px;padding-top:12px;font-size:14px;line-height:1.55}
#detail .body a{color:#58a6ff;cursor:pointer}#detail .body pre{background:var(--bg);padding:10px;border-radius:6px;overflow:auto}
#detail .body table{border-collapse:collapse}#detail .body td,#detail .body th{border:1px solid var(--border);padding:4px 8px}
.backlinks a,.title-link{color:#58a6ff;cursor:pointer;display:block;padding:2px 0;font-size:13px}
.hint{color:var(--mut);font-size:13px}h3.sec{font-size:12px;text-transform:uppercase;letter-spacing:.05em;color:var(--mut);margin:16px 0 4px}
#legend{padding:8px 12px;border-top:1px solid var(--border);font-size:12px;color:var(--mut);display:flex;flex-wrap:wrap;gap:10px}
#legend span{display:inline-flex;align-items:center;gap:4px}#legend i{width:10px;height:10px;border-radius:50%;display:inline-block}
</style></head><body>
<div id="app">
<div id="graph"></div>
<div id="side">
<div id="controls">
<input id="search" placeholder="Search title, id, tags…">
<select id="typeFilter"><option value="">all types</option></select>
<select id="layout">
<option value="cose">cose</option><option value="concentric">concentric</option>
<option value="breadthfirst">breadth-first</option><option value="circle">circle</option><option value="grid">grid</option>
</select>
</div>
<div id="detail"><p class="hint">Click a node to inspect it. Drag to pan, scroll to zoom.</p></div>
<div id="legend"></div>
</div>
</div>
<script>
const G={"bundle": "knowledge", "nodes": [{"id": "concepts/compounding_artifact", "path": "concepts/compounding_artifact.md", "type": "Concept", "title": "Compounding Artifact", "description": "The defining property of an LLM Wiki — knowledge is compiled once and kept current, so the wiki gets richer with every source and every query.", "tags": ["pattern", "core"], "resource": "", "status": "active", "body": "# Compounding Artifact
The key difference between an [LLM Wiki](./llm_wiki.md) and query-time retrieval is that
**the wiki is a persistent, compounding artifact**. The cross-references are already there. The
contradictions have already been flagged. The synthesis already reflects everything you've read.
The wiki keeps getting richer with every source you add and every question you ask.
## Two ways knowledge compounds
**Ingested sources compound.** Each new source is not merely indexed for later retrieval; it is
integrated. A single source might touch 10–15 pages — updating entity pages, revising summaries,
adding cross-links. See [ingest](../operations/ingest.md).
**Queries compound too.** A good answer — a comparison, a multi-source analysis, a discovered
connection — should not disappear into chat history. It can be filed back into the wiki as a new
page. This way explorations accumulate just like sources do. See [query](../operations/query.md).
## Contrast
This is precisely what [RAG](./rag_vs_llm_wiki.md) does *not* do: with retrieval, the
LLM rediscovers knowledge from scratch on every question and nothing is built up. The compounding
property is also why the [lint](../operations/lint.md) operation matters — a compounding artifact
accumulates drift (stale claims, orphans, contradictions) that must be periodically swept.
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md)", "links": ["concepts/llm_wiki", "operations/ingest", "operations/query", "concepts/rag_vs_llm_wiki", "operations/lint", "references/karpathy_llm_wiki"], "cited_by": ["concepts/llm_wiki", "concepts/rag_vs_llm_wiki", "ecosystem/omegawiki", "operations/ingest", "operations/lint", "operations/query", "references/karpathy_llm_wiki", "references/okf_vs_rag_infographic"]}, {"id": "concepts/concept_document", "path": "concepts/concept_document.md", "type": "Concept", "title": "Concept Document", "description": "The OKF term for a single unit of knowledge — a UTF-8 markdown file with a YAML frontmatter block and a markdown body.", "tags": ["okf", "vocabulary"], "resource": "", "status": "active", "body": "# Concept Document
A **concept document** (or *concept*) is a single unit of knowledge in a
[knowledge bundle](./knowledge_bundle.md): a UTF-8 markdown file with a
[YAML frontmatter block](../spec/frontmatter.md) followed by a [markdown body](../spec/body.md).
Every `.md` file in a bundle is a concept document *except* the reserved
[`index.md` and `log.md`](../spec/reserved_filenames.md).
Every file you are reading in this bundle's `concepts/`, `spec/`, `operations/`,
`implementations/`, and `references/` directories is a concept document.
## Concept ID
A concept's identifier is its file path within the bundle with the `.md` extension removed.
For example, this file's concept ID is `concepts/concept_document`. Concept IDs are how concepts
are referenced and how [cross-links](../spec/cross_linking.md) resolve.
## Anatomy
* **Frontmatter** — machine-readable metadata. Only [`type`](../spec/frontmatter.md) is required;
`title`, `description`, `resource`, `tags`, and `timestamp` are recommended.
* **Body** — human- and agent-readable markdown prose, with conventional headings such as
`# Schema`, `# Examples`, and `# Citations` when applicable. See [Body](../spec/body.md).
A concept may be **bound to a resource** (e.g. a database table, with a `resource` URI) or
**abstract** (e.g. a playbook or, as here, an idea) with no `resource`. Every concept in this
bundle is abstract.
# Citations
1. [OKF Specification (SPEC.md)](../references/okf_spec.md)", "links": ["concepts/knowledge_bundle", "spec/frontmatter", "spec/body", "spec/reserved_filenames", "spec/cross_linking", "references/okf_spec"], "cited_by": ["concepts/knowledge_bundle", "concepts/three_layer_architecture", "operations/ingest", "operations/lint", "references/qmd", "spec/citations", "spec/cross_linking", "spec/frontmatter", "spec/index_files", "spec/motivation", "spec/reserved_filenames", "spec/terminology"]}, {"id": "concepts/knowledge_bundle", "path": "concepts/knowledge_bundle.md", "type": "Concept", "title": "Knowledge Bundle", "description": "The OKF term for a directory tree of concept documents — the unit of production, exchange, and consumption.", "tags": ["okf", "vocabulary"], "resource": "", "status": "active", "body": "# Knowledge Bundle
A **knowledge bundle** (or just *bundle*) is the OKF unit of knowledge: a directory tree of
markdown [concept documents](./concept_document.md) plus optional reserved
[index](../spec/index_files.md) and [log](../spec/log_files.md) files. It is the OKF formalization
of the \"wiki\" layer in the [three-layer architecture](./three_layer_architecture.md).
See [Bundle Structure](../spec/bundle_structure.md) for the normative rules. In short:
* A bundle is just a directory — no manifest or database is required.
* It can be distributed as a git repository (recommended), an archive (tar/zip), or a
subdirectory of a larger repository. This repository uses the last form: the bundle lives in
`knowledge/`.
* Its internal organization is domain-independent; producers arrange concepts as they see fit.
* It may declare its format version via `okf_version` in the root
[index file](../spec/index_files.md) — see [Versioning](../spec/versioning.md).
Because a bundle is plain markdown and YAML, anyone can produce one (people, agents on any
framework, export pipelines) and anyone can consume one (file servers, Obsidian/Notion, LLMs,
search indexes, graph viewers). This vendor-neutrality is the whole point of
[OKF](../spec/index.md).
# Citations
1. [OKF Specification (SPEC.md)](../references/okf_spec.md)", "links": ["concepts/concept_document", "spec/index_files", "spec/log_files", "concepts/three_layer_architecture", "spec/bundle_structure", "spec/versioning", "references/okf_spec"], "cited_by": ["concepts/concept_document", "concepts/llm_wiki", "concepts/progressive_disclosure", "concepts/three_layer_architecture", "ecosystem/kiso", "ecosystem/openknowledge_cli", "implementations/okf_native_agent", "operations/query", "references/okf_readme", "references/qmd", "spec/bundle_structure", "spec/terminology"]}, {"id": "concepts/llm_wiki", "path": "concepts/llm_wiki.md", "type": "Concept", "title": "LLM Wiki", "description": "A persistent, LLM-maintained knowledge base that compiles knowledge once and keeps it current, sitting between you and your raw sources.", "tags": ["pattern", "knowledge-base", "core"], "resource": "", "status": "active", "body": "# LLM Wiki
The **LLM Wiki** is a pattern for building knowledge bases where an LLM agent incrementally
builds, cross-references, and maintains a structured, interlinked collection of markdown files
that sits between you and your raw sources. The pattern originates from
[Andrej Karpathy's idea file](../references/karpathy_llm_wiki.md).
The core move: instead of retrieving from raw documents at query time
(see [RAG vs. LLM Wiki](./rag_vs_llm_wiki.md)), the LLM reads each new source, extracts
the key information, and integrates it into the existing wiki — updating entity pages, revising
summaries, flagging contradictions, strengthening the evolving synthesis. Knowledge is
**compiled once and then kept current**, not re-derived on every query. This is what makes the
wiki a [compounding artifact](./compounding_artifact.md).
## Division of labor
You never (or rarely) write the wiki yourself. The LLM owns the writing and maintenance; you
own sourcing, exploration, and asking good questions. This split is what makes the pattern
practical — see [why it works](#why-it-works).
The workflow Karpathy describes: the LLM agent open on one side, a markdown editor
(e.g. Obsidian) on the other. The LLM edits based on the conversation; you browse the results
in real time. *\"Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.\"*
## Structure
An LLM Wiki has [three layers](./three_layer_architecture.md): immutable raw sources,
the LLM-owned wiki, and a schema/config document that tells the LLM how the wiki is organized.
The wiki itself can be expressed as an OKF [knowledge bundle](./knowledge_bundle.md) —
that is exactly what this repository does.
## Operations
Three operations keep the wiki alive:
* [Ingest](../operations/ingest.md) — process a new source into the wiki.
* [Query](../operations/query.md) — answer a question, and file good answers back.
* [Lint](../operations/lint.md) — periodically health-check the wiki.
## Why it works
The tedious part of maintaining a knowledge base is not the reading or the thinking — it is the
bookkeeping: updating cross-references, keeping summaries current, noting contradictions,
maintaining consistency across dozens of pages. Humans abandon wikis because the maintenance
burden grows faster than the value. LLMs don't get bored, don't forget to update a
cross-reference, and can touch fifteen files in one pass. The wiki stays maintained because the
cost of maintenance is near zero. This is the same problem the [Memex](./memex.md) could
not solve in 1945: who does the maintenance.
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md)", "links": ["references/karpathy_llm_wiki", "concepts/rag_vs_llm_wiki", "concepts/compounding_artifact", "concepts/three_layer_architecture", "concepts/knowledge_bundle", "operations/ingest", "operations/query", "operations/lint", "concepts/memex"], "cited_by": ["concepts/compounding_artifact", "concepts/memex", "concepts/rag_vs_llm_wiki", "concepts/three_layer_architecture", "design/sample_bundle_lessons", "ecosystem/critiques", "ecosystem/openwiki_langchain", "implementations/personal_work_wiki", "operations/ingest", "references/karpathy_llm_wiki", "spec/cross_linking", "spec/log_files", "spec/motivation", "spec/versioning"]}, {"id": "concepts/memex", "path": "concepts/memex.md", "type": "Concept", "title": "Memex", "description": "Vannevar Bush's 1945 vision of a personal, curated knowledge store with associative trails — the intellectual antecedent of the LLM Wiki, blocked only by the maintenance problem.", "tags": ["lineage", "history"], "resource": "", "status": "active", "body": "# Memex
The [LLM Wiki](./llm_wiki.md) is related in spirit to Vannevar Bush's **Memex**,
described in his 1945 essay *As We May Think*. The Memex was a vision of a personal, curated
knowledge store with **associative trails** between documents — where the connections between
documents are as valuable as the documents themselves.
Bush's vision was closer to the LLM Wiki than to what the web actually became: private, actively
curated, with links as first-class value. The web optimized for scale and publication; the Memex
imagined a private thinking tool.
## The part Bush couldn't solve
The Memex assumed a human would build and maintain the trails. That is exactly the
[bookkeeping burden](./llm_wiki.md#why-it-works) that causes humans to abandon knowledge
bases: maintaining associative links by hand doesn't scale, and the value decays as the store
grows. The LLM handles that maintenance — it doesn't get bored and can update every
cross-reference in one pass. In this framing, the LLM Wiki is the Memex with the maintenance
problem finally solved.
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md)
2. Vannevar Bush, \"As We May Think,\" *The Atlantic*, July 1945.", "links": ["concepts/llm_wiki", "references/karpathy_llm_wiki"], "cited_by": ["concepts/llm_wiki", "ecosystem/critiques", "references/karpathy_llm_wiki"]}, {"id": "concepts/progressive_disclosure", "path": "concepts/progressive_disclosure.md", "type": "Concept", "title": "Progressive Disclosure", "description": "Navigating a growing bundle one level at a time through index files, so an agent finds relevant pages without loading everything.", "tags": ["pattern", "navigation"], "resource": "", "status": "active", "body": "# Progressive Disclosure
**Progressive disclosure** is the technique of navigating a [knowledge bundle](./knowledge_bundle.md)
one directory level at a time, using [index files](../spec/index_files.md) as the map, rather than
loading the whole bundle into context. It is how both agents and humans stay oriented as a wiki
grows to hundreds of pages.
## How it works
Each directory can carry an `index.md` that lists its contents — a link and one-line description
per concept, grouped under headings. An agent answering a [query](../operations/query.md) reads the
root index first, follows the relevant section index, and only then reads the specific concept
pages it needs. The [cross-links](../spec/cross_linking.md) between concepts let it fan out from
there. This works surprisingly well at moderate scale (~100 sources, hundreds of pages) and
avoids the need for embedding-based retrieval infrastructure.
## Why it scales
Index files are cheap to read and give the agent a content-oriented catalog of what exists.
Because the bundle is [graph-shaped, not just tree-shaped](../spec/cross_linking.md), the agent can
also traverse relationships directly once it has a foothold. When a bundle outgrows what index
files can handle, add a search tool such as [qmd](../references/qmd.md) — but not before.
Progressive disclosure is a primary motivation for keeping [index files](../spec/index_files.md)
current on every [ingest](../operations/ingest.md); a stale index defeats the mechanism.
# Citations
1. [OKF README & Reference Agent](../references/okf_readme.md)
2. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md)", "links": ["concepts/knowledge_bundle", "spec/index_files", "operations/query", "spec/cross_linking", "references/qmd", "operations/ingest", "references/okf_readme", "references/karpathy_llm_wiki"], "cited_by": ["concepts/rag_vs_llm_wiki", "design/skill_design", "ecosystem/critiques", "operations/ingest", "operations/query", "references/okf_vs_rag_infographic", "references/qmd", "spec/bundle_structure", "spec/index_files"]}, {"id": "concepts/rag_vs_llm_wiki", "path": "concepts/rag_vs_llm_wiki.md", "type": "Concept", "title": "RAG vs. LLM Wiki", "description": "Retrieval-augmented generation rediscovers knowledge on every query; an LLM Wiki compiles it once and maintains it.", "tags": ["comparison", "rag"], "resource": "", "status": "active", "body": "# RAG vs. LLM Wiki
Most people's experience with LLMs and documents is **RAG** (retrieval-augmented generation):
you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates
an answer. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.
The [LLM Wiki](./llm_wiki.md) pattern is different in one decisive way.
| | RAG | LLM Wiki |
|---|---|---|
| When knowledge is synthesized | At query time, from raw chunks | At ingest time, into pages |
| Persistence | Nothing accumulates between queries | A [compounding artifact](./compounding_artifact.md) |
| Cross-references | Re-discovered per question | Already written and maintained |
| Contradictions | Re-encountered each time | Flagged once, during ingest |
| Answer to a 5-document question | Re-piece the fragments every time | Read the page that already synthesized them |
| Infrastructure | Embeddings + vector store | Markdown files + [index](../spec/index_files.md) (add search later) |
RAG's weakness is not accuracy but **amnesia**: the LLM is rediscovering knowledge from scratch
on every question, so nothing is built up. Ask a subtle question that requires synthesizing five
documents and the LLM has to find and piece together the fragments every single time.
The LLM Wiki front-loads that synthesis into the [ingest](../operations/ingest.md) step, so at
query time the work is mostly navigation ([progressive disclosure](./progressive_disclosure.md))
rather than rediscovery. The two are not mutually exclusive — a large wiki can still use search
such as [qmd](../references/qmd.md) to find pages — but the retrieved unit is a synthesized page,
not a raw chunk.
For a visual version of this contrast — \"search all 100 PDFs every query\" vs. \"read once, compile
one concept per file, follow links\" — see the
[OKF vs. RAG infographic](../references/okf_vs_rag_infographic.md).
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md)", "links": ["concepts/llm_wiki", "concepts/compounding_artifact", "spec/index_files", "operations/ingest", "concepts/progressive_disclosure", "references/qmd", "references/okf_vs_rag_infographic", "references/karpathy_llm_wiki"], "cited_by": ["concepts/compounding_artifact", "concepts/llm_wiki", "operations/query", "references/karpathy_llm_wiki", "references/okf_vs_rag_infographic", "references/qmd"]}, {"id": "concepts/three_layer_architecture", "path": "concepts/three_layer_architecture.md", "type": "Concept", "title": "Three-Layer Architecture", "description": "An LLM Wiki has three layers — immutable raw sources, the LLM-owned wiki, and a co-evolved schema/config document.", "tags": ["pattern", "architecture", "core"], "resource": "", "status": "active", "body": "# Three-Layer Architecture
The [LLM Wiki](./llm_wiki.md) pattern separates concerns into three layers. Keeping them
distinct is what makes the LLM a disciplined maintainer rather than a generic chatbot.
## 1. Raw sources
Your curated collection of source documents: articles, papers, images, transcripts, data files.
These are **immutable** — the LLM reads from them but never modifies them. This is your source
of truth and your audit trail. In the [ingest](../operations/ingest.md) operation, a source is
read once and then moved to a \"processed\" location; it is never rewritten.
## 2. The wiki
A directory of LLM-generated markdown files: summaries, entity pages, concept pages,
comparisons, overviews, syntheses. **The LLM owns this layer entirely** — it creates pages,
updates them as new sources arrive, maintains cross-references, and keeps everything consistent.
You read it; the LLM writes it. Expressed in OKF, this layer is a
[knowledge bundle](./knowledge_bundle.md) of [concept documents](./concept_document.md).
## 3. The schema
A configuration document — `CLAUDE.md`, `AGENTS.md`, or an in-bundle config — that tells the LLM
how the wiki is structured, what the conventions are, and what workflows to follow when
ingesting, querying, or maintaining. Karpathy calls this \"the key configuration file.\" It is
co-evolved over time as you learn what works for your domain.
This layer is the pivot for **portability**: if the operations are generic and all
domain knowledge (the section taxonomy, the `type` values, the workflows) lives in the schema,
then the same skills can drive a work wiki, a book companion, or a research corpus. The portable
skills we are building read this schema; the schema makes them fit the project.
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md)", "links": ["concepts/llm_wiki", "operations/ingest", "concepts/knowledge_bundle", "concepts/concept_document", "references/karpathy_llm_wiki"], "cited_by": ["concepts/knowledge_bundle", "concepts/llm_wiki", "design/skill_design", "design/spec_evolution", "ecosystem/competitor_comparison", "ecosystem/okf_skills_scaccogatto", "ecosystem/openknowledge_cli", "ecosystem/wiki_skills", "implementations/personal_work_wiki", "operations/ingest", "references/karpathy_llm_wiki", "references/okf_spec"]}, {"id": "design/sample_bundle_lessons", "path": "design/sample_bundle_lessons.md", "type": "Concept", "title": "Lessons from the OKF Sample Bundles", "description": "Our analysis of the three reference OKF bundles (ga4, stackoverflow, crypto_bitcoin) — what their structure, frontmatter, and linking conventions teach us about authoring conformant bundles.", "tags": ["okf", "examples", "analysis", "design-input"], "resource": "", "status": "active", "body": "# OKF Sample Bundles
The OKF repo ships three reference bundles produced by the
[reference agent](../references/okf_readme.md): **ga4** (17 files), **stackoverflow** (53 files), and
**crypto_bitcoin** (8 files). All three describe BigQuery public datasets. Reading them is the best
way to see the [spec](../spec/index.md) as *practiced* rather than *stated* — and several of their
conventions differ from choices we made in this bundle, which is useful design input.
## Shared structure
All three use the same shallow, domain-driven layout:
```
<bundle>/
index.md # root: a \"# Subdirectories\" list, NO frontmatter, NO okf_version
datasets/ # the dataset concept + index.md
tables/ # one concept per table + index.md
references/ # (ga4, stackoverflow) enums, metrics, joins, licenses + index.md
viz.html # generated visualizer output, checked in
```
Only two `type` values do the heavy lifting: **`BigQuery Table`** (resource-bound) and
**`Reference`** (abstract). Consistent with the spec's advice to pick few, descriptive types.
## What they confirm
* **Reserved files carry no frontmatter.** Every `index.md` is just a markdown list; none declare
`okf_version` — confirming that declaration is genuinely optional (we chose to include it).
* **Resource-bound concepts** (`tables/*.md`, `datasets/*.md`) set a real `resource` URI (the
BigQuery REST endpoint) and lead with a `# Schema` section — the §4.3 pattern.
* **Abstract concepts** (`references/*.md` — enums like `post_type_ids`, metrics, joins) omit
`resource` and just carry prose + a SQL snippet — the §4.4 pattern.
* **`# Citations`** appears at the bottom as a bulleted list of URLs (sometimes plain, sometimes
markdown links) — looser than a numbered list, still conformant.
* **`timestamp`** is a full ISO-8601 datetime (`'2026-05-28T22:53:05+00:00'`), quoted — worth
matching for machine-friendliness (we used date-only).
## What surprised me (and what it teaches)
1. **Root `index.md` is minimal — a `# Subdirectories` list, not a full catalog.** Progressive
disclosure is done by *drilling into each directory's* `index.md`, not by one big root index.
Our root index is much richer. Both conform; theirs scales better to large bundles, ours is
friendlier to a human reader landing at the top. A portable skill should probably generate the
minimal per-directory style by default.
2. **No bundle has a `log.md`.** The reference agent produces bundles as a *snapshot* from a source
of truth (BigQuery), so there's no change history to keep. This confirms `log.md` matters for the
*incrementally-maintained* [LLM Wiki](../concepts/llm_wiki.md) use case (our case,
[personal work wiki](../implementations/personal_work_wiki.md), [the OKF-native agent](../implementations/okf_native_agent.md)) but is
genuinely optional for generated catalogs.
3. **They use relative links (`../references/metrics/event_count.md`), not the recommended
bundle-absolute (`/…`) form.** The [spec](../spec/cross_linking.md) *recommends* absolute; the
reference bundles use relative. A reminder that \"recommended\" ≠ \"required,\" and that relative
links are what actually renders on GitHub and in the visualizer. (We chose absolute for stability;
worth revisiting when we decide what the skills emit.)
4. **The `events_` table concept is enormous** — one file with the full nested BigQuery schema
(hundreds of fields). OKF does not force atomicity; a concept can be as large as its resource. This
contrasts with [the OKF-native agent's](../implementations/okf_native_agent.md) \"prefer atomic concepts\"
guidance — atomicity is a *maintenance* choice for living wikis, not an OKF rule.
## Implications for our skills
* Default to **per-directory `index.md`** catalogs (progressive disclosure) rather than one giant
root index.
* Make **`log.md` conditional** on the use case: incremental wiki → yes; generated snapshot → no.
* Decide a **link-form policy** (absolute vs. relative) and apply it consistently; the ecosystem is
split, so pick one and document why.
* Use **full ISO-8601 datetimes** for `timestamp`.
* Don't over-enforce **atomicity** at the format layer — make it a schema-layer/style preference.
# Citations
1. OKF sample bundles — <https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf/bundles>
2. [OKF README & Reference Agent](../references/okf_readme.md)", "links": ["references/okf_readme", "concepts/llm_wiki", "implementations/personal_work_wiki", "implementations/okf_native_agent", "spec/cross_linking"], "cited_by": ["design/spec_evolution", "ecosystem/kiso", "references/okf_vs_rag_infographic", "spec/cross_linking"]}, {"id": "design/skill_design", "path": "design/skill_design.md", "type": "Concept", "title": "Skill Design — the kb-* family", "description": "The plan for agent-knowledge's skills — a kb hub (router + single source of truth) plus action skills (ingest, query, init, lint, visualize), designed with the writing-great-skills framework.", "tags": ["design", "skills", "build-plan", "decisions"], "resource": "", "status": "active", "body": "# Skill Design — the `kb-*` family
The build plan for this project's skills, designed against the
[competitor comparison](../ecosystem/competitor_comparison.md) (what to build vs. reuse) and the
`writing-great-skills` framework (predictability as the root virtue). Naming is `kb-` — \"knowledge
bundle\" is OKF's own term ([SPEC §2](../references/okf_spec.md)) for the `knowledge/` package the
skills operate on, and `kb` reads universally as \"knowledge base.\" Nothing user-facing says \"okf\";
OKF-conformance is carried *internally*.
## The family
A **hub** skill plus action skills, on the `git`/`kb-<verb>` model:
| Skill | Invocation | Role |
|---|---|---|
| **`kb`** | model-invoked | Router + single source of truth. Teaches the system and its terms; routes to the right action skill; holds the shared spec, glossary, trust-model, templates, and example bundle. |
| **`kb-init`** | user-invoked | Scaffold a bundle from the hub's templates/example. Default `knowledge/`; custom path allowed; multi-bundle aware. |
| **`kb-ingest`** | model-invoked | Raw source → integrated concepts, applying the trust model. The core differentiator. |
| **`kb-query`** | model-invoked | Progressive-disclosure discovery → synthesize with citations → file good answers back. |
| **`kb-lint`** | user-invoked | Drift + health checks; delegates the deterministic §9 conformance pass to an existing validator. |
| **`kb-visualize`** | user-invoked | LLM-generated graph — native UI where supported, self-contained HTML artifact otherwise. |
| **`kb-search`** *(later)* | user-invoked | qmd/FTS retrieval when a bundle outgrows [index-style discovery](../concepts/progressive_disclosure.md). |
## Framework decisions
### Invocation split — context load vs. cognitive load
A model-invoked skill's **description** sits in the context window every turn (context load); a
user-invoked skill costs nothing until typed (but spends the human's cognitive load). Rule: model-
invoke only what must fire autonomously. So the **daily verbs** [`kb-ingest`](../operations/ingest.md)
and [`kb-query`](../operations/query.md) — which must fire on \"capture this\" / \"what do we know
about X\" — are model-invoked, plus the `kb` hub as the front door. The **deliberate** skills
(`kb-init`, `kb-lint`, `kb-visualize`, `kb-search`) are user-invoked. Net: **three descriptions in
context, not seven.**
### `kb` as router + single source of truth
Two framework moves converge on the hub. A **router** cures the cognitive load of several user-
invoked siblings by naming them and when to reach each. A **single source of truth** keeps one
authoritative copy of shared meaning. So `kb` holds, once, what every action skill needs:
```
kb/
SKILL.md # teach the system + key terms; route to kb-<verb>
reference/
SPEC.md # OKF v0.1, vendored verbatim — the source of truth
glossary.md # bundle, concept, reference, progressive disclosure, compounding…
trust-model.md # append-only / supersede / conflicts_with rules
templates/
concept.md index.md log.md
example-bundle/ # tiny canonical bundle: teaching example AND kb-init's seed
```
Action skills point here rather than restating the spec — avoiding the drift the OKF reference
implementation hit when the spec was copied into multiple modules (PR #161). A spec bump to 0.2 is
then a one-file edit.
### Information hierarchy
Each action `SKILL.md` carries ordered **steps** with **completion criteria** that are *checkable*
and *exhaustive*; shared **reference** (spec, glossary, trust rules) is disclosed to `kb/reference/`
behind context pointers (progressive disclosure) so each skill stays legible.
### Leading words
Reused across skills, docs, and this bundle's own concept names so the shared language sharpens both
execution and invocation: **bundle**, **ingest**/**compile**, **supersede** (the trust model's
verb), **drift** (what [lint](../operations/lint.md) fights), **progressive disclosure**,
**compounding**.
### Guarded failure mode — premature completion in `kb-ingest`
[Ingest](../operations/ingest.md) is the long sequence (read → extract → integrate across concepts →
update indexes → append [log](../spec/log_files.md) → move source to processed → commit); the tail
tempts an early \"done.\" Defense, in order: sharpen the completion criteria first (*\"every extracted
entity has a concept or is consciously skipped; source moved to processed; log appended\"*); split the
sequence only if the rush persists.
### Techniques to borrow for `kb-ingest` (from openwiki)
LangChain's [openwiki](../ecosystem/openwiki_langchain.md) is in the adjacent code-docs lane (not
OKF), but its agent prompt is proven and directly applicable to `kb-ingest`:
- **Ground every claim** in an inspected source / git evidence — never invent. Operationalizes the
trust model's \"never lose provenance.\"
- **Plan-then-write** — draft a temporary plan (intended concepts touched + evidence + open
questions) before writing, so discovery precedes synthesis. A natural front-half for the ingest
sequence and a guard against the premature-completion failure above.
- **Read-only research subagents** — 1-2 narrow-brief subagents *inspect and summarize only*; the
main agent does all writes. Lets a large source set be processed in parallel without write
conflicts, and answers the token-cost critique.
## Build vs. reuse (from the competitor comparison)
**Build** (differentiated): `kb-init` (schema-layer scaffold), `kb-ingest` (trust-modeled raw-source
loop — nobody has this), `kb-query` (light), `kb-visualize` (adaptive, LLM-generated — not a fixed
`viz.html`). **Reuse:** the deterministic §9 validator inside `kb-lint` rather than rebuilding it.
See [competitor comparison](../ecosystem/competitor_comparison.md).
## Decisions (settled 2026-07-01)
1. **Distribution → skills.sh, single-plugin layout** (mattpocock/skills model). Each skill is a
folder `skills/<name>/SKILL.md` with supporting files beside it; a `.claude-plugin/plugin.json`
lists them; no build step. Installs via skills.sh *and* works as a Claude Code plugin. The whole
`kb-*` family **ships as one plugin** so the action skills can reach the [`kb`](#the-family) hub's
shared `reference/` (vendoring the spec into each skill would reintroduce drift — rejected).
2. **Schema-layer location → in-bundle `spec/` concepts.** `kb-init` writes a bundle's domain rules
(concept `type` vocabulary, folder taxonomy, ingest conventions) as real OKF concepts under
`knowledge/<bundle>/spec/` — the pattern Google's cricket bundle (#144) and *this* bundle already
use. Portable, stays [conformant](../spec/conformance.md), travels with the bundle, and the rules
are themselves browsable/linkable knowledge. This is the [schema
layer](../concepts/three_layer_architecture.md#3-the-schema) made concrete and is what lets the
generic skills fit any domain.
3. **Trust model → full [the OKF-native agent](../implementations/okf_native_agent.md) model, opinionated.**
Append-only on meaning, supersede-with-provenance (`status`/`supersedes`/`superseded_by`),
`conflicts_with` over silent overwrite, events-additive. An opinionated default *is* the
predictability the framework prizes — not a per-run configuration choice. Answers the ecosystem's
top [critique](../ecosystem/critiques.md#1-truth-maintenance-and-knowledge-base-poisoning).
## Build order
`kb` (hub + shared reference) → `kb-init` (proves scaffold + schema layer) → `kb-ingest` (the
differentiator) → `kb-query` → `kb-lint` → `kb-visualize`. `kb-search` deferred until a bundle needs
it.
# Citations
1. `writing-great-skills` framework (user-invoked skill, consulted 2026-07-01).
2. [Competitor Comparison](../ecosystem/competitor_comparison.md)
3. [OKF Spec Evolution](./spec_evolution.md)", "links": ["ecosystem/competitor_comparison", "references/okf_spec", "concepts/progressive_disclosure", "operations/ingest", "operations/query", "operations/lint", "spec/log_files", "ecosystem/openwiki_langchain", "spec/conformance", "concepts/three_layer_architecture", "implementations/okf_native_agent", "ecosystem/critiques", "design/spec_evolution"], "cited_by": ["ecosystem/landscape"]}, {"id": "design/spec_evolution", "path": "design/spec_evolution.md", "type": "Concept", "title": "OKF Spec Evolution (open PRs & proposals)", "description": "Our reading of the open pull requests and proposals on GoogleCloudPlatform/knowledge-catalog — where OKF v0.1 is still in flux (link form, required frontmatter, an emerging trust/provenance axis) and what it means for our build.", "tags": ["okf", "spec", "evolution", "design-input", "analysis"], "resource": "", "status": "active", "body": "# OKF Spec Evolution
Reading the [open PRs](https://github.com/GoogleCloudPlatform/knowledge-catalog/pulls) on the OKF
repo is the best signal for **what in v0.1 is still moving** — and several threads land directly on
decisions this project must make. Captured as of 2026-07-01; treat as a snapshot, re-check before
finalizing skills.
## 1. Link form is being *reversed* (recommend relative, not absolute)
The single most decision-relevant thread. The [spec](../spec/cross_linking.md) currently *recommends*
absolute bundle-relative (`/…`) links, but the reference agent's prompt **forbids** them (\"never
start a link with `/` — that breaks GitHub rendering\") and every shipped
[sample bundle](./sample_bundle_lessons.md) uses **relative** links. PRs **#165**, **#58**,
**#66**, **#110**, **#161** all converge on fixing the spec to match practice:
* **#165** swaps §5.1/§5.2 so **relative links become \"recommended\"** (they resolve in any renderer —
`cat`, browser, GitHub, editor — with no OKF tooling), and reframes absolute as \"requires an
OKF-aware resolver.\"
* The failure mode: a leading `/` resolves to the **repository/host root**, not the bundle root,
whenever a bundle ships as a subdirectory (§3 explicitly allows this) — exactly our layout
(`knowledge/` inside a larger repo).
**Implication for us — done.** We originally chose *absolute* links throughout this bundle. On
2026-07-01 we **converted the whole bundle to relative links** (473 links across 51 files) so it
renders correctly on GitHub and for any non-OKF-aware reader, aligning with the reference agent, the
sample bundles, and #165's direction. Note the *spec-text* PR is unlikely to merge soon — as of this
writing #165/#66/#110 are all open with **no maintainer review, only CLA-bot activity** — but the
decision doesn't depend on it: relative is already the de-facto-correct form for a nested bundle.
See the [cross-linking spec page](../spec/cross_linking.md).
## 2. Only `type` is required — confirmed and being enforced
PRs **#145**, **#64**, **#161** fix the reference implementation, which wrongly treated `title` /
`description` / `timestamp` as REQUIRED and would reject spec-minimal bundles. The spec is
authoritative: **only [`type`](../spec/frontmatter.md) is required**; the four-key check is a
producer *quality bar*, not conformance. This **validates our
[conformance](../spec/conformance.md) checker** (type-only) exactly.
## 3. A trust / provenance / reliability axis is emerging
Multiple independent efforts are converging on the very problem we identified as our
[differentiator](../ecosystem/competitor_comparison.md) and the ecosystem's top
[critique](../ecosystem/critiques.md). This is the thread to watch most closely:
* **#58 (§12 Trust & safety)** — consumers MUST treat bundle contents as untrusted **data, never
instructions** (prompt-injection); OKF gives no built-in authenticity guarantee. Also reconciles
the §6↔§11 frontmatter contradiction and adds README/LICENSE/CONTRIBUTING as ignored files.
* **#159 (`reliability`)** — an optional frontmatter convention for *epistemic* reliability: a
maturity ladder (`confidence` + `basis` → corroboration tiers), with honesty rules like
\"signed ≠ verified\" and \"`verified` requires ≥2 sources.\"
* **#50 (`sources`)** — optional machine-readable provenance alongside the human `# Citations`
section (which source systems produced this, can it be refreshed/audited, content digest).
* Related issues #92/#94 (groundedness), #140 (integrity/signing), #99 (policy receipts).
**Implication for us.** Our planned [trust model](../ecosystem/critiques.md) (append-only, supersede,
`conflicts_with`) is squarely aligned with where OKF itself is heading — good. But the format may
**standardize the field names** (`reliability`, `sources`, `confidence`). We should adopt *their*
emerging names as [extension keys](../spec/frontmatter.md#extensions) rather than invent our own, and
treat ingested source content as untrusted data (a real prompt-injection surface for
[ingest](../operations/ingest.md)).
## 4. Reserved-file handling is tightening
**#149** fixes the reference impl to exclude `log.md` (not just `index.md`) from concept
enumeration — it had been showing up as an `Unknown` node. **#58** proposes treating
`README.md`/`LICENSE.md`/`CONTRIBUTING.md` as ignored non-concepts so a plain README doesn't make a
bundle non-conformant. Our [checker](../spec/conformance.md) already handles `index.md`/`log.md`; we
should add the README/LICENSE tolerance when we harden it.
## 5. Domain (hand-authored) bundles are being legitimized
**#144** adds a hand-authored **cricket** bundle (vs. the DB-generated samples), with per-bundle
`spec/types.md`, `spec/provenance.md`, and `spec/sample-size.md` files, a novel `story` type, and
additive keys (`source_boundary`, `entity_id`, `same_as`). This directly validates two of our
bets: (a) **domain-knowledge bundles** are first-class OKF (our case, not just data catalogs), and
(b) **putting the taxonomy/conventions in per-bundle `spec/` files** is a real pattern — essentially
the [schema layer](../concepts/three_layer_architecture.md#3-the-schema) we want to formalize, and
which this very bundle already uses (`/spec`).
## 6. Keeping source material *in* the bundle (§8 + discussion #91)
An easy-to-miss clause of SPEC **§8**: citation links MAY point into a **`references/` subdirectory
that mirrors external material as first-class OKF concepts** — i.e. the spec explicitly blesses
storing PDFs, images, transcripts, and `.mov` files *inside* the bundle, wrapped as
`type: Reference` concepts, rather than only linking out to URLs.
Discussion **#91** (opened by this project's author) resolves the \"where do I keep source docs\"
question, with the community answer: **separate the canonical source from the derived text** — keep
a stable pointer to the original asset in the bundle *and* carry extracted text/summary for
retrieval; keep genuinely incidental material external. See our [citations](../spec/citations.md)
page, which now documents this, and the [infographic](../references/okf_vs_rag_infographic.md) reference that
applies it.
## Resolved decisions for this bundle
* **Link form → relative (done 2026-07-01).** Converted all 473 links across 51 files from absolute
(`/…`) to relative, per §1 above.
* **Trust-model field names → adopt upstream (pending build).** When we build the skills, use the
emerging OKF names (`reliability`, `sources`, `confidence`, `status`/`supersedes`/`superseded_by`/
`conflicts_with`) as [extension keys](../spec/frontmatter.md#extensions) rather than inventing our
own, and treat ingested source content as untrusted data.
* **Reserved/ignored files → widen the checker (pending build).** Add README/LICENSE/CONTRIBUTING
tolerance (per #58) alongside the existing `index.md`/`log.md` handling.
# Citations
1. OKF pull requests — <https://github.com/GoogleCloudPlatform/knowledge-catalog/pulls>
2. PR #165 (relative links), #145/#64/#161 (required frontmatter), #58 (trust/safety), #159 (reliability), #50 (sources), #149 (reserved files), #144 (cricket domain bundle)", "links": ["spec/cross_linking", "design/sample_bundle_lessons", "spec/frontmatter", "spec/conformance", "ecosystem/competitor_comparison", "ecosystem/critiques", "operations/ingest", "concepts/three_layer_architecture", "spec/citations", "references/okf_vs_rag_infographic"], "cited_by": ["design/skill_design", "spec/cross_linking"]}, {"id": "ecosystem/commonplace", "path": "ecosystem/commonplace.md", "type": "Reference", "title": "commonplace (zby)", "description": "A theory-forward, review-gated framework for agent-operated knowledge — typed, linked, review-gated markdown that agents execute — and a maintained list of related systems (~70 stars).", "tags": ["ecosystem", "featured", "review-gated", "theory"], "resource": "https://github.com/zby/commonplace", "status": "active", "body": "# commonplace (zby)
`zby/commonplace` (~70★) describes itself as *\"the theory of LLM wikis, running as one\"* — a
framework for agent-operated knowledge that is **typed, linked, and review-gated** markdown the
agents execute. Homepage: <https://zby.github.io/commonplace/>.
## Why it's worth studying
* **Review-gated** is the direct embodiment of the top [critique's](./critiques.md#1-truth-maintenance-and-knowledge-base-poisoning)
proposed fix — the agent proposes, a gate approves — rather than \"the LLM owns the layer entirely.\"
* **Typed + linked** aligns closely with OKF's [`type`](../spec/frontmatter.md) +
[cross-link](../spec/cross_linking.md) model; a good reference for how far to push typing.
* It maintains an **agent-curated index of related systems**
(<https://zby.github.io/commonplace/notes/related-systems/related-systems-index/>) — itself a live
example of a wiki used to survey its own [ecosystem](./landscape.md), and a useful
external map to cross-check ours against.
## Relevance to us
commonplace is the most *theory-forward* project in the thread and the clearest articulation of the
review-gated stance. When we decide the default autonomy level of our [ingest](../operations/ingest.md)
skill (fully autonomous vs. propose-and-approve), this is the reference argument for the cautious
end — complementary to [the OKF-native agent's](../implementations/okf_native_agent.md) provenance-based trust
model.
# Citations
1. commonplace — <https://github.com/zby/commonplace>
2. commonplace related-systems index — <https://zby.github.io/commonplace/notes/related-systems/related-systems-index/>", "links": ["ecosystem/critiques", "spec/frontmatter", "spec/cross_linking", "ecosystem/landscape", "operations/ingest", "implementations/okf_native_agent"], "cited_by": []}, {"id": "ecosystem/competitor_comparison", "path": "ecosystem/competitor_comparison.md", "type": "Concept", "title": "Competitor Comparison — okf-skills vs. openknowledge vs. our plan", "description": "A feature-by-feature comparison of the two most direct OKF-native competitors against what this project should build, isolating the differentiated surface.", "tags": ["ecosystem", "competitor", "design-input", "strategy"], "resource": "", "status": "active", "body": "# Competitor Comparison
A feature-by-feature read of the two most direct [OKF](../spec/index.md)-native competitors —
[okf-skills (scaccogatto)](./okf_skills_scaccogatto.md) and
[openknowledge (openknowledge-sh)](./openknowledge_cli.md) — against what *we* plan to
build, based on reading their actual skill files, templates, and validation matrices (not just their
READMEs). Goal: **build only the differentiated parts.**
## At a glance
| | **okf-skills** | **openknowledge** | **ours (planned)** |
|---|---|---|---|
| Form factor | Claude Code plugin + skills.sh skills | Standalone Go CLI (+ Codex skill) | Portable agent skills |
| Runtime / deps | Agent + `uv`/python for validator | Native Go binary (installer) | Agent-only, no runtime dep |
| Cross-agent | ✅ 20+ agents via skills.sh | Agent-agnostic (CLI); ships a Codex skill | ✅ (target Claude Code first) |
| Author/produce | ✅ `produce` mode | ✅ `new` + agent `setup` prompt | ◐ via ingest |
| **Maintain-in-sync** | ✅ `maintain` mode (code/docs changed) | ◐ maintenance-loop guidance in skill | ✅ **core** |
| **Ingest from raw sources** | ✗ (authors from code/docs) | ✗ (scaffolds, human edits md) | ✅ **core differentiator** |
| Consume/query | ✅ `consume` mode | ✅ `use` (prints entrypoint/excerpt) | ✅ |
| Deterministic validation | ✅ `okf_validate.py` (§9) | ✅ Go validator + compliance matrix + tests | ◐ reuse, don't rebuild |
| Visualize | ✅ `viz.html` (Cytoscape) | ✅ `to html/json/graph/tar` exporters | ✗ (reuse theirs) |
| Multi-bundle / registry | ✗ | ✅ **registry** (local/published/archive/Git) | ◐ (OKF-native-agent-style multi-kb) |
| Publish to web | ✗ | ✅ static site + `llms.txt` + manifest | ✗ (hand off to [kiso](./kiso.md)/openknowledge) |
| **Trust model** (append-only / supersede / conflicts) | ✗ (`**Deprecation**` note only) | ✗ (changelog page) | ✅ **core differentiator** |
| **Domain portability seam** (schema layer) | ◐ (pick a layout by domain) | ◐ (`setup` tailors to use case) | ✅ **explicit** |
| Spec pinning | ✅ vendored verbatim SPEC.md | ✅ embedded, version-selectable | ✅ (vendor verbatim) |
| Dogfoods itself in OKF | ✅ `.okf/` + CI validation | ✅ `Wiki/` + decisions/workflows | ✅ `knowledge/` (this bundle) |
Legend: ✅ strong / ◐ partial / ✗ absent.
## What both already do well (don't rebuild)
* **Deterministic §9 [conformance](../spec/conformance.md) validation.** okf-skills ships a
self-contained `okf_validate.py`; openknowledge has a Go validator with a full hard-rule
compliance matrix backed by tests. Reinventing this is pure waste — we should **reuse
okf-skills' validator** (MIT, `uv run`, zero-config) as our [lint](../operations/lint.md)'s
conformance pass and spend our effort on the *drift* checks it doesn't do.
* **Visualization / export.** Both render a bundle to a self-contained graph; openknowledge also
exports json/tar/graph and a static site. We should **not** build a visualizer — point users at
`okf:visualize` or `openknowledge to html`.
* **Spec-pinning discipline.** Both vendor the spec verbatim as the skill's source of truth. We do
this too (`references/okf_spec.md` points at it); worth vendoring the literal file into any skill
we ship.
## Where the competitors are thin (our opening)
1. **Ingest from raw, messy sources.** Both are *authoring* tools: okf-skills `produce`s concepts
from **code/docs/manual** input; openknowledge `new` scaffolds an empty bundle a human then
edits. Neither has the [LLM-Wiki ingest loop](../operations/ingest.md) — drop a transcript / email /
PDF / screenshot, extract entities and signals, integrate across many concepts, move the source to
processed. This is the [personal work wiki](../implementations/personal_work_wiki.md) workflow and our clearest differentiator.
2. **A real trust / truth-maintenance model.** okf-skills' maintain mode says \"update the body and
`timestamp`, add a `**Deprecation**` note\" — it *edits claims in place*. Neither implements
[the OKF-native agent's](../implementations/okf_native_agent.md) append-only-on-meaning, supersede-with-
provenance, `conflicts_with`, events-are-additive model — which is exactly the fix the
[ecosystem's top critique](./critiques.md#1-truth-maintenance-and-knowledge-base-poisoning)
demands. Making that the *default* maintenance behavior is a genuine, defensible difference.
3. **The [schema-layer](../concepts/three_layer_architecture.md#3-the-schema) portability seam.** Both
\"pick a layout by domain,\" but the domain knowledge is improvised per-run. Neither has a
first-class, per-project config that declares the taxonomy, `type` vocabulary, and workflow so the
*same* generic skills fit a work wiki, a book companion, or a research corpus with no skill edits.
## What to borrow outright
* **Dual distribution** (plugin + skills.sh, scripts via `${CLAUDE_SKILL_DIR}`) — okf-skills' answer
to \"ship portable skills.\" Adopt wholesale.
* **`setup` prints an agent prompt** (openknowledge) — deterministic scaffold + agent judgment for
the use-case-specific parts. A great shape for our `init` skill.
* **`use` prints an entrypoint** (openknowledge) — a path-light way for an agent to load the right
knowledge on demand; better than hardcoding `index.md` reads.
* **Positioning table** (okf-skills) — OKF vs. `CLAUDE.md` vs. auto-memory vs. wiki. Adopt for our docs.
* **CI-validated self-dogfooding** (both) — add a CI conformance check on `knowledge/`.
* **Agent-maintenance footer** `<!-- okf-footer: agent-maintenance -->` (openknowledge) — a tidy
convention for keeping source-anchors/update-notes out of prominent headings. Worth stealing for
our concept template.
* **Delegate bounded maintenance to focused low-reasoning subagents** (openknowledge) — a concrete
answer to the [token-cost critique](./critiques.md#2-token-cost-is-postponed-not-eliminated).
## Recommended scope for our skills
Build **three** skills, thin where the field is saturated and thick where it's empty:
* **`okf-init`** — scaffold `knowledge/` + a per-project [schema-layer](../concepts/three_layer_architecture.md#3-the-schema)
config; borrow openknowledge's print-a-prompt setup. Vendor the spec.
* **`okf-ingest`** — the differentiated core: raw source → integrated concepts, with the
[trust model](./critiques.md) (append-only / supersede / conflicts) as default. Wraps the
[ingest](../operations/ingest.md) + [query](../operations/query.md) operations.
* **`okf-lint`** — drift + health checks ([lint](../operations/lint.md)), delegating the deterministic
§9 pass to okf-skills' validator rather than reimplementing it.
Do **not** build: a visualizer, an exporter/publisher, or another from-scratch conformance checker —
reuse okf-skills and openknowledge/[kiso](./kiso.md) for those.
# Citations
1. [okf-skills (scaccogatto)](./okf_skills_scaccogatto.md) — SKILL.md files, templates, validator
2. [openknowledge (openknowledge-sh)](./openknowledge_cli.md) — skill, tooling-model & spec-compliance docs
3. [Critiques & Open Problems](./critiques.md)", "links": ["ecosystem/okf_skills_scaccogatto", "ecosystem/openknowledge_cli", "ecosystem/kiso", "spec/conformance", "operations/lint", "operations/ingest", "implementations/personal_work_wiki", "implementations/okf_native_agent", "ecosystem/critiques", "concepts/three_layer_architecture", "operations/query"], "cited_by": ["design/skill_design", "design/spec_evolution"]}, {"id": "ecosystem/critiques", "path": "ecosystem/critiques.md", "type": "Concept", "title": "Critiques & Open Problems", "description": "The substantive objections to the LLM Wiki pattern from the community — truth maintenance / KB poisoning, token cost at scale, and markdown-vs-database — and what they imply for our design.", "tags": ["ecosystem", "critique", "design-input"], "resource": "", "status": "active", "body": "# Critiques & Open Problems
The most useful signal in the [gist](../references/karpathy_llm_wiki.md) thread is not the flood of
implementations but the recurring, well-argued **objections**. Each maps to a concrete design
choice for our portable skills.
## 1. Truth maintenance and knowledge-base poisoning
The single most-repeated concern (commenter `laphilosophia`; the \"hidden flaw\" article by
`foundanand`). The appealing part of the pattern — the LLM owns synthesis — is exactly where models
fail *quietly*: bad synthesis, stale claims surviving new evidence, page sprawl, and **false
consistency** accumulate invisibly because every output still looks coherent. The sharpest framing:
once LLM-authored summaries are indexed alongside sources, later passes retrieve and reason over
*AI output* rather than ground truth — a self-referential drift away from reality.
The risky sentence is *\"the LLM owns this layer entirely.\"* Fine for low-stakes personal use; too
aggressive for team or high-accuracy contexts.
**Proposed fixes (from the thread):** source-grounded, citation-first, **review-gated** wikis where
the LLM proposes patches rather than being the final authority; extract only *verifiable structure*
(entities, relationships, citations) and keep narrative synthesis lighter or query-time; tie every
claim to a source, an uncertainty level, and recency.
**Implication for us:** this is the strongest argument for adopting
[the OKF-native agent's trust model](../implementations/okf_native_agent.md) — append-only on meaning,
supersede-with-provenance, `conflicts_with` over silent overwrite, and mandatory
[citations](../spec/citations.md) — as the *default* behavior of our [ingest](../operations/ingest.md)
and [lint](../operations/lint.md) skills, not an optional extra. It also raises the stakes on
[lint](../operations/lint.md): drift detection is a core feature, not hygiene.
## 2. Token cost is postponed, not eliminated
(`jgravelle`, \"a radical diet…\"; echoed by `YokoPunk`, `druce`, and others.) The pattern trades
per-query retrieval cost for per-session **compilation** cost. That holds until the wiki outgrows
the context window: past roughly 50–100K tokens the [index.md](../spec/index_files.md) becomes a
bottleneck, [progressive-disclosure](../concepts/progressive_disclosure.md) navigation gets
unreliable, and answers degrade. *\"You didn't eliminate retrieval. You postponed it.\"*
**Proposed fixes:** treat the wiki as a queryable dataset — `search_sections` / `get_section`
retrieval of only relevant chunks (claims of ~95% context reduction); add TLDRs at the top of each
page so an index scan → TLDR → drill-down saves tokens; add a real search engine.
**Implication for us:** [qmd](../references/qmd.md) (or an FTS/section-retrieval tool) is not a
\"someday\" nicety — it is the answer to the scaling objection, and our [query](../operations/query.md)
skill should be built to *degrade gracefully* from pure index navigation to search-backed retrieval.
Consider a `TLDR`/`description`-first read convention.
## 3. Markdown vs. database
A principled minority (`buremba`/owletto in PostgreSQL; `gnusupport`'s \"Hyperscope\" PostgreSQL
argument) hold that for **deterministic**, strongly-typed knowledge with reliable queries, a
database with an event log beats a pile of markdown. Entity types get strict schemas; the agent
gets SQL.
**Implication for us:** this is a genuine trade-off, not a mistake to refute. Markdown wins on
portability, diffability, human-readability, and zero-infra — the whole [OKF thesis](../spec/motivation.md).
A database wins on query determinism and typed integrity. OKF's [frontmatter](../spec/frontmatter.md)
is the hedge: structured, queryable metadata on top of prose. Worth stating explicitly in our docs
*when* the file-based approach is the wrong tool.
## 4. \"Isn't this just…?\"
Recurring skeptical takes worth answering, not dismissing: *\"just structured context / a good
`AGENTS.md` hierarchy\"* (`skpalan`), *\"NotebookLM already does this,\"* *\"this is the
[Zettelkasten](https://zettelkasten.de/introduction/) / second-brain idea again,\"* and the
[Memex](../concepts/memex.md)/Engelbart lineage. The honest answer: the *novel* part is not the wiki
but **who maintains it** — the pattern is only interesting because the LLM makes the bookkeeping
cost near-zero (see [why it works](../concepts/llm_wiki.md#why-it-works)). The lint pass — periodic
self-audit — is what most \"just structured context\" setups lack and what everyone concedes is
valuable.
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md) (comments: laphilosophia, skpalan, YokoPunk, buremba, gnusupport)
2. \"The Hidden Flaw in Karpathy's LLM Wiki\" — <https://foundanand.medium.com/the-hidden-flaw-in-karpathys-llm-wiki-e3a86a94b459>
3. \"A Radical Diet for Karpathy's Token-Eating LLM Wiki\" — <https://dev.to/jgravelle/a-radical-diet-for-karpathys-token-eating-llm-wiki-59ng>", "links": ["references/karpathy_llm_wiki", "implementations/okf_native_agent", "spec/citations", "operations/ingest", "operations/lint", "spec/index_files", "concepts/progressive_disclosure", "references/qmd", "operations/query", "spec/motivation", "spec/frontmatter", "concepts/memex", "concepts/llm_wiki"], "cited_by": ["design/skill_design", "design/spec_evolution", "ecosystem/commonplace", "ecosystem/competitor_comparison", "ecosystem/karpathy_llm_wiki_astro", "ecosystem/landscape", "ecosystem/okf_skills_scaccogatto", "ecosystem/openknowledge_cli", "ecosystem/openwiki_langchain", "ecosystem/synthadoc", "ecosystem/wiki_skills", "references/okf_vs_rag_infographic"]}, {"id": "ecosystem/karpathy_llm_wiki_astro", "path": "ecosystem/karpathy_llm_wiki_astro.md", "type": "Reference", "title": "karpathy-llm-wiki (Astro-Han)", "description": "An Agent-Skills-compatible LLM wiki for Claude Code, Cursor, and Codex — build a Karpathy-style knowledge base from raw sources with citations and linting (~1.3k stars).", "tags": ["ecosystem", "featured", "skills", "multi-agent-tool"], "resource": "https://github.com/Astro-Han/karpathy-llm-wiki", "status": "active", "body": "# karpathy-llm-wiki (Astro-Han)
`Astro-Han/karpathy-llm-wiki` (~1.3k★) is an **Agent-Skills-compatible** LLM wiki that works across
Claude Code, Cursor, and Codex. It builds a Karpathy-style knowledge base from raw sources with
first-class **citations and linting**.
## Why it's worth studying
* It is the closest high-traction analogue to **our exact deliverable**: portable *skills* (not a
platform) that make an agent maintain a wiki, and explicitly multi-tool rather than Claude-only.
* Citations and [lint](../operations/lint.md) are built in — an implicit acknowledgment of the
[truth-maintenance critique](./critiques.md#1-truth-maintenance-and-knowledge-base-poisoning).
## Relevance to us
The key differentiator remains **format**: like almost all of the [ecosystem](./landscape.md),
this project uses a bespoke/Obsidian-flavored convention rather than a conformant
[OKF](../spec/index.md) bundle. It is the strongest prior art for \"how to package the workflow as
cross-agent skills,\" and the best place to study skill ergonomics before we write our own. Compare
directly with [wiki-skills](./wiki_skills.md).
# Citations
1. karpathy-llm-wiki — <https://github.com/Astro-Han/karpathy-llm-wiki>", "links": ["operations/lint", "ecosystem/critiques", "ecosystem/landscape", "ecosystem/wiki_skills"], "cited_by": ["ecosystem/landscape", "ecosystem/wiki_skills"]}, {"id": "ecosystem/kiso", "path": "ecosystem/kiso.md", "type": "Reference", "title": "kiso (oak-invest)", "description": "A Java publishing engine that turns OKF bundles into static websites for humans and AI agents, emitting llms.txt and sitemap.xml. An OKF consumer, not a producer.", "tags": ["ecosystem", "featured", "okf-native", "publishing", "consumer"], "resource": "https://github.com/oak-invest/kiso", "status": "active", "body": "# kiso (oak-invest)
`oak-invest/kiso` (~11★) is *\"a publishing engine that turns
[Open Knowledge Format](../spec/index.md) bundles into static websites for humans and AI agents.\"*
Written in Java, distributed as a native-image CLI and a GitHub Action.
## What it is (and isn't)
kiso sits on the **consumer** side of OKF, whereas okf-skills, the OKF-native agent, and this project sit
on the **producer/maintainer** side. It reads a conformant [bundle](../concepts/knowledge_bundle.md)
and generates a static site:
```bash
kiso-cli build --source=examples/kb-google-example --destination=public
```
Notably it emits **`llms.txt` and `sitemap.xml`** alongside the HTML — i.e. it publishes the same
bundle for *both* human browsers and AI agents. Its example bundle (`kb-google-example`) is a
near-copy of the official [GA4 sample](../design/sample_bundle_lessons.md), confirming those bundles
are becoming the ecosystem's de-facto interop test.
## Why it matters to us
* **It validates the \"produce once, consume many\" thesis** at the heart of
[OKF](../spec/motivation.md): kiso is proof that a bundle we produce with our skills can be consumed
by a completely independent tool (a different language, a different vendor) with no coordination —
the whole point of adopting a [spec](../spec/index.md) over a bespoke convention.
* **`llms.txt` + `sitemap.xml` output** is a concrete idea for making a published bundle
agent-consumable on the open web — a possible downstream target for bundles our skills maintain.
* Because it consumes the [GA4 sample bundle](../design/sample_bundle_lessons.md) directly, kiso is a
ready-made **conformance smoke test**: if kiso can build our bundle into a site, we're
interoperable.
## Relevance to us
kiso is not a competitor — it is the consumer half of the ecosystem our producer skills feed. It
strengthens the case for strict [conformance](../spec/conformance.md): the more independent consumers
like kiso exist, the more valuable it is that our output is genuinely spec-conformant rather than
\"markdown that mostly works.\" Worth keeping as a downstream target and interop check.
# Citations
1. kiso — <https://github.com/oak-invest/kiso>", "links": ["concepts/knowledge_bundle", "design/sample_bundle_lessons", "spec/motivation", "spec/conformance"], "cited_by": ["ecosystem/competitor_comparison", "ecosystem/landscape", "ecosystem/openknowledge_cli", "ecosystem/openwiki_langchain"]}, {"id": "ecosystem/landscape", "path": "ecosystem/landscape.md", "type": "Concept", "title": "LLM Wiki Ecosystem Landscape", "description": "A survey of the 200+ implementations spawned by Karpathy's gist, grouped by shape — skills/plugins, CLIs, apps/platforms, MCP servers, and databases-not-markdown.", "tags": ["ecosystem", "survey"], "resource": "", "status": "active", "body": "# LLM Wiki Ecosystem Landscape
Within weeks of the [gist](../references/karpathy_llm_wiki.md), its comment thread became a directory
of 200+ implementations. They cluster into a few recognizable shapes. Understanding the clusters
tells us where our [OKF-conformant, portable-skills](../implementations/index.md) angle is
differentiated and where it is crowded.
## 1. Agent skills & plugins
Packages of `SKILL.md` / plugin files that turn Claude Code, Cursor, Codex, etc. into a wiki
maintainer — the shape closest to **what we are building**.
* [karpathy-llm-wiki (Astro-Han)](./karpathy_llm_wiki_astro.md) — Agent-Skills-compatible, multi-tool.
* [wiki-skills (kfchou)](./wiki_skills.md) — Claude Code skills, Karpathy-faithful.
* Others: `doneyli/claude-code-plugins` (llm-wiki), `horiacristescu/claude-playbook-plugin`,
`EveryInc/compound-engineering-plugin` (22k★, broader \"compound engineering\"),
`pedronauck/skills` (karpathy-kb), `Thrimbda/legion-mind`, `vanillaflava/llm-wiki-claude-skills`,
`FBoschman/claude-wiki-research-skills`, `theafh/ai-modules`.
This category is busy but almost entirely **Obsidian-wikilink-flavored**, not
[OKF](../spec/index.md)-conformant — our differentiator.
## 2. CLIs & compilers
Standalone tools that ingest sources and compile a wiki, often with built-in search.
* [synthadoc](./synthadoc.md) — no-tools, self-managed compiler (534★).
* `Hosuke/llmbase` — ingest→compile→query→enhance with a React UI (42★).
* `VihariKanukollu/browzy.ai` — npm CLI, FTS5+BM25, Obsidian-compatible, multi-model.
* Others: `doum1004/llmwiki-cli`, `olegiv/llm-wiki-go` (Go codebase wiki),
`atomicmemory/llm-wiki-compiler`, `iamsashank09/llm-wiki-kit`, `MauricioPerera/llm-wiki-kit`,
`yhay81/create-wiki-kit`.
## 3. Apps & platforms
Full products with UIs, hosting, or mobile.
* [OmegaWiki](./omegawiki.md) — the most complete (1.5k★).
* `Tencent/WeKnora` (17k★) — RAG + agent + self-maintaining wiki (enterprise-scale, RAG-centric).
* `memex-lab/memex` (568★) — local-first AI journal app for iOS/Android.
* `basicmachines-co/basic-memory` (3.3k★) — \"conversations that remember,\" MCP-based.
* `bitsofchris/openaugi` (119★), `sheawinkler/contextlattice` (122★, coordination control-plane).
## 4. MCP servers & memory layers
Expose the wiki/memory to any agent over MCP rather than shipping a workflow.
* `multimail-dev/thinking-mcp`, `Electro-resonance/LLM-WIKI-MCP`, `deepak-bhardwaj-ps/smriti-mcp`,
`gowtham0992/link` (link-mcp), `dfalci/mcp-advwiki`, `us/crw` (research/ingest via MCP).
## 5. Coordination & multi-agent
Let several agents build one wiki in parallel.
* [tracecraft (Arrmlet)](./landscape.md) — shared memory/messaging/task-claiming over any
S3 bucket (27★); `swarmclawai/swarmvault`, `redmizt/multi-agent-wiki-toolkit`,
`AEVYRA/llm-wiki-coordination`.
## 6. Databases, not markdown
The notable dissent from the file-based approach — see [critiques](./critiques.md).
* `buremba` / `lobu-ai/owletto` — entity-typed knowledge in **PostgreSQL** with an event log and SQL
access for the agent.
* `gnusupport`'s \"Hyperscope\" argument for PostgreSQL over markdown for *deterministic* KBs.
* `zTgx/vectorless` / `vectorlessflow` — \"knowing by reasoning, not vectors.\"
## 7. Codebase-doc generators (adjacent lane)
Auto-document a *codebase* for agents — generated from the repo, not ingested from arbitrary
sources. Overlaps in spirit (agent-maintained, `AGENTS.md` pointer, scheduled updates) but is a
different product than an OKF general-knowledge bundle.
* [openwiki (langchain-ai)](./openwiki_langchain.md) — LangChain/DeepAgents CLI; daily-PR updates.
Not OKF, no trust model — but a strong category signal and a source of
[ingest techniques](../design/skill_design.md). Peers: DeepWiki, `garrytan/gbrain`.
## Where OKF sits
A handful of projects target [OKF](../spec/index.md) explicitly, and the cohort is growing fast:
* [okf-skills (scaccogatto)](./okf_skills_scaccogatto.md) — Claude Code skills to author/validate/visualize bundles (**our closest competitor**).
* [openknowledge (openknowledge-sh)](./openknowledge_cli.md) — a Go CLI with a registry, viewer, exporters, and agent maintenance loop (the most tooling-complete).
* [kiso (oak-invest)](./kiso.md) — a Java engine that publishes bundles as static sites (the *consumer* side).
* [okf-harness (pumblus)](./okf_harness.md) — an agent-first local harness for OKF wikis.
* `equationalapplications/expo-llm-wiki` — cites the OKF SPEC; SQLite-backed memory.
* Our sibling [the OKF-native agent](../implementations/okf_native_agent.md) — OKF-native deployable agent.
The vast majority of the ecosystem still reinvents a bespoke on-disk convention. But the OKF-native
cohort is now real and moving fast — [okf-skills](./okf_skills_scaccogatto.md) (skills),
[openknowledge](./openknowledge_cli.md) (CLI + registry), and [kiso](./kiso.md)
(publisher) already bracket authoring, tooling, and consumption. The remaining gap our project aims
at: **portable skills that incrementally produce and *maintain* a conformant bundle from raw
sources, with a [trust model](./critiques.md)** — the living-wiki
[ingest](../operations/ingest.md) loop, which the authoring/validation/publishing tools largely leave
to the human. A caution worth noting: with several teams shipping OKF tooling in mid-2026, our value
has to be the *maintenance loop and portability seam*, not another author/validate/visualize trio.
# Citations
1. [Karpathy — LLM Wiki gist](../references/karpathy_llm_wiki.md) (comment thread, ~900 comments)", "links": ["references/karpathy_llm_wiki", "ecosystem/karpathy_llm_wiki_astro", "ecosystem/wiki_skills", "ecosystem/synthadoc", "ecosystem/omegawiki", "ecosystem/critiques", "ecosystem/openwiki_langchain", "design/skill_design", "ecosystem/okf_skills_scaccogatto", "ecosystem/openknowledge_cli", "ecosystem/kiso", "ecosystem/okf_harness", "implementations/okf_native_agent", "operations/ingest"], "cited_by": ["ecosystem/commonplace", "ecosystem/karpathy_llm_wiki_astro", "ecosystem/okf_harness", "ecosystem/okf_skills_scaccogatto"]}, {"id": "ecosystem/okf_harness", "path": "ecosystem/okf_harness.md", "type": "Reference", "title": "okf-harness (pumblus)", "description": "An agent-first local harness for OKF-compatible LLM Wikis — one of the few community projects targeting the OKF format explicitly (~17 stars).", "tags": ["ecosystem", "featured", "okf-native", "harness"], "resource": "https://github.com/pumblus/okf-harness", "status": "active", "body": "# okf-harness (pumblus)
`pumblus/okf-harness` (~17★) is an *\"agent-first local harness for OKF-compatible LLM Wikis.\"* Stars
aside, it is significant because it is one of the **very few** community projects in the
[ecosystem](./landscape.md) that targets [OKF](../spec/index.md) by name rather than a
bespoke convention.
## Why it's worth studying
* It shares our core bet: **OKF-conformant bundles as the substrate**, agent as the operator.
* \"Local harness\" framing is adjacent to our \"portable skills\" framing — worth comparing how it
structures [ingest](../operations/ingest.md)/[query](../operations/query.md) around a conformant
bundle, and whether it stays strictly conformant or extends the format.
* Together with `equationalapplications/expo-llm-wiki` (which cites the OKF SPEC) and our sibling
[the OKF-native agent](../implementations/okf_native_agent.md), it marks the small but real OKF-native cohort.
## Relevance to us
This is the most direct \"someone else is doing OKF\" data point. Before we finalize skill design,
it's worth reading okf-harness to avoid reinventing conventions and to see where a second
implementer felt the [spec](../spec/index.md) needed extending (a strong signal for where OKF v0.1 is
thin — e.g. relationship/state vocabulary, which [the OKF-native agent](../implementations/okf_native_agent.md)
also had to add).
# Citations
1. okf-harness — <https://github.com/pumblus/okf-harness>", "links": ["ecosystem/landscape", "operations/ingest", "operations/query", "implementations/okf_native_agent"], "cited_by": ["ecosystem/landscape"]}, {"id": "ecosystem/okf_skills_scaccogatto", "path": "ecosystem/okf_skills_scaccogatto.md", "type": "Reference", "title": "okf-skills (scaccogatto)", "description": "A Claude Code-native OKF toolchain — okf/validate/visualize skills plus a plugin and skills.sh distribution, driven by the verbatim spec with a deterministic conformance checker. The closest direct competitor to this project.", "tags": ["ecosystem", "featured", "okf-native", "competitor", "claude-code", "skills"], "resource": "https://github.com/scaccogatto/okf-skills", "status": "active", "body": "# okf-skills (scaccogatto)
`scaccogatto/okf-skills` (~21★) is *\"the OKF toolkit for Claude Code — author, maintain, validate &
visualize Open Knowledge Format bundles.\"* It is the **closest direct competitor** to what this
project set out to build: portable, OKF-conformant agent skills. Studying it carefully is worth
more than any other entry in the [ecosystem](./landscape.md).
## What it ships