Possible Paths: Self-Updating Wiki and LLM-Wiki Integration

Date: 2026-05-23
Status: Exploratory / logged for future reference
Context: Recent research into Karpathy’s LLM Wiki gist and the Hermes Agent bundled research/llm-wiki skill, alongside ongoing work on the root workbench/ synthesis system and the 4-tier (L4 → L3 → L2 → L1) pipeline.

The Core Question

The current knowledge base is deliberately human-curated at the high-signal end:

• Raw material enters via raw/inbox/, raw/sources/, and sessions.
• High-value topics receive multi-model L3 drafting in workbench/.
• L2 applies strict voice and integrity standards (mechanism-first, no artifact-as-subject, artifact-as-subject avoided, density, provenance awareness).
• Finished L2s promote to wiki/ under Dimensions, Concepts, Syntheses, Workflows, Systems, etc.

This produces excellent, readable, Five-Dimensions-aligned pages. The cost is friction and limited compounding on the long tail of smaller observations, clippings, and recurring patterns.

The question is whether (and how) to introduce elements of a self-updating, compounding wiki — where the base itself becomes an active participant that extracts, cross-references, maintains consistency, and surfaces gaps — without eroding the voice quality or the priority on deliberate human synthesis.

The Reference Pattern (Karpathy + Hermes Implementation)

Karpathy’s LLM Wiki gist describes a persistent markdown knowledge base that the LLM incrementally builds and maintains instead of treating every query as a fresh retrieval problem. The wiki is the compiled knowledge.

The Hermes Agent research/llm-wiki skill (version 2.1+) implements this with a concrete three-layer architecture that has proven practical:

• Layer 1 – Raw (immutable): raw/articles/, raw/papers/, raw/transcripts/. Sources are captured once with frontmatter (source_url, ingested, sha256). They are never edited.
• Layer 2 – Wiki pages: entities/, concepts/, comparisons/, queries/. These are the living, agent-updated pages with YAML frontmatter (title, created/updated, type, tags, sources, confidence, contested, contradictions).
• Layer 3 – Schema and navigation: SCHEMA.md (domain conventions, tag taxonomy, page thresholds, update policy), index.md (sectioned catalog with one-line summaries), log.md (append-only action history).

Critical rituals and mechanisms:

• Orientation on every session: Always read SCHEMA + index + recent log before touching content. This prevents duplicates and contradictions.
• Provenance: ^[raw/...] markers on synthesized claims; sources: frontmatter.
• Page thresholds and splitting rules: Only create pages when something meets centrality or multi-source criteria; split at ~200 lines.
• Lint as a first-class operation: Orphan detection, broken wikilinks, frontmatter validation, staleness, contradiction surfacing, tag audit, log rotation.
• Obsidian-native: Wikilinks, Dataview, graph view all work immediately.
• Division of labor: Human curates sources and direction; agent handles extraction, cross-referencing, filing, and consistency maintenance.

This produces a wiki that gets smarter over time with relatively low per-ingest human cost for routine material.

How This Relates to the Existing System

The current setup already has pieces that point in similar directions:

• wiki/Workflows/Raw to Wiki Compilation.md and Wiki Breakdown Pass.md, Wiki Health Checks.md describe ingestion, backlinking, and maintenance.
• There are existing index.md patterns in various folders and a Source Index.
• The workbench/ + normalized model-prefixed files + index.md per topic is already a strong “active synthesis cockpit”.
• Strong emphasis on mechanism-first language and living style rules (style-feedback.md) that any automation must respect.
• Recent Claude Cowork synthesis work surfaces exactly the capability layers (skills, scheduled tasks, autonomous builders, mission control) that could drive periodic wiki maintenance.

The gaps relative to the Karpathy/Hermes pattern are mainly:

• No single top-level SCHEMA.md with explicit tag taxonomy and page-creation thresholds for the whole base.
• Limited automatic provenance tracking across the long tail.
• Cross-referencing and contradiction surfacing are manual or semi-manual.
• The “long tail” raw material does not reliably become lightweight wiki entities/concepts.
• No equivalent to the “always orient first” ritual that keeps an agent from drifting when the wiki grows large.

Possible Paths (Exploratory Options)

These are not mutually exclusive and can be staged. They are logged here as possibilities to revisit when friction or scale makes the trade-offs clearer.

Path 1 — Schema + Provenance Hardening (Lowest friction, highest compatibility)

• Add a root-level wiki/SCHEMA.md (or per major area) that codifies current conventions plus new rules drawn from the Karpathy pattern: file naming, frontmatter fields (including confidence, contested, contradictions), tag taxonomy aligned with Five Dimensions + Concepts, page thresholds, update policy, provenance marker syntax.
• Introduce lightweight raw/wiki-ingest/ or mirror selected raw sources with sha256 frontmatter.
• Update the existing Raw to Wiki Compilation workflow and the L3→L2 converter to emit provenance markers and required frontmatter fields.
• Enhance index.md generation (possibly via a skill) so every new L2 promotion automatically appears in the right section.
• Add a recurring “wiki lint” pass (manual at first, later skill- or script-driven) that surfaces orphans, broken links, low-confidence pages, and contradictions.

This path strengthens what already exists without changing the human-heavy high-signal synthesis loop.

Path 2 — Dedicated Wiki Compiler Skill + Workbench Handoff

• Build (or evolve the previous attempt at) a synthesis-workbench skill that understands the current root workbench/ layout, normalized filenames, and topic index.md.
• Add a companion or extension skill: wiki-ingest or wiki-compiler that can:
  • Take a finished L2 (or a batch of raw clippings).
  • Perform entity/concept extraction against the current wiki SCHEMA.
  • Propose new pages or updates with draft wikilinks and provenance.
  • Update navigation files.
  • Generate a lint report for human review.
• The workbench remains the place for deliberate voice-polished synthesis. The compiler skill handles the “compile the long tail and maintain the graph” work.
• Trigger phrases could include “ingest this into the wiki”, “run wiki compilation on recent inbox”, “lint the knowledge base”.

This creates a clean division: workbench for high-craft L3/L2 work; wiki compiler for compounding and hygiene.

Path 3 — Parallel Lightweight LLM-Wiki Layer (More ambitious)

• Maintain the current curated wiki/ as the “reviewed, voice-aligned, Five-Dimensions-spined” layer.
• Introduce a second directory (e.g., wiki-llm/ or data/wiki-raw/) that follows the Hermes/Karpathy three-layer structure more strictly.
• New raw material can flow into the LLM-wiki layer with lower ceremony.
• Periodic (or skill-triggered) “promotion” passes move high-value, well-supported entities/concepts into the main curated wiki, with human voice editing.
• This gives the best of both worlds: fast compounding on volume + protected high-signal core.
• Requires clearer boundaries and possibly different Obsidian vault settings or symlinks.

Path 4 — Hybrid with External Agent (Hermes or similar)

• Continue primary work in the current Grok Build + Obsidian + Quartz environment.
• Periodically hand off batches or the entire raw/ + workbench/ output to a Hermes session (which already has the excellent llm-wiki skill and supports Grok models).
• Use Hermes for heavy lifting on cross-referencing, lint, and entity extraction, then pull the resulting markdown back into the main vault for voice review and promotion.
• This leverages the fact that Hermes already ships the exact skill and has strong Grok integration.

Path 5 — Skills-First Evolution (Meta-path)

Regardless of which architectural path is chosen, the highest-leverage near-term move is probably skills that reduce the activation energy of the current pipeline while making future wiki-style behaviors cheap:

• Correctly scoped synthesis-workbench skill for the new root layout (create topic folders, normalize files, generate comparison index.md, archive completed topics).
• Upgraded light-voice-evolution skill that embeds the latest artifact-as-subject, negative-framing, and Pre-Write Integrity Pass rules.
• New skills: wiki-promoter (L2 → L1 with backlink and index hygiene), entity-extractor, contradiction-surfacer, blueprint-extractor (neutral mechanism extraction from ambitious workflow posts like the recent Claude Cowork sources).
• Scheduled or on-demand tasks that run lint + daily/weekly wiki briefings back into the journal or a “Knowledge Base as Thinking Partner” page.

These skills can be built incrementally and will pay dividends on any of the paths above.

Tensions and Guardrails

• Voice quality vs. scale: Any automation must pass through or be reviewable against the living style rules. The Pre-Write Integrity Pass and Light Voice Evolution process are non-negotiable for anything that reaches L2 or promoted wiki pages.
• Human agency: The knowledge base is a thinking partner, not a replacement. Automatic updates should surface candidates and contradictions for human resolution, not silently harden claims.
• Five Dimensions spine: New structures must continue to make the Dimensions (Mindset, Self-Management, Self-Regulation, Deep Processing, Retrieval) the primary navigation and meaning-making frame.
• Reversibility: Any new layer or schema should be easy to unwind or migrate if it increases friction instead of reducing it.
• Obsidian + Quartz dual surface: Everything must remain pleasant to read and navigate in both the local vault and the public static site.

Relation to Existing Thinking

This exploration directly extends:

• Knowledge Base as Thinking Partner — the wiki becomes a more active, compounding second brain while keeping the first brain in the loop.
• Workflows page and the various Wiki Health / Compilation workflows.
• The Claude Cowork as Personal Work OS synthesis (especially the ambitious blueprint layer around skills, scheduled autonomous work, mission control, and self-documenting systems).
• The recent directory and pipeline cleanup that made workbench/ the single cockpit for active L3/L2 work.

Concrete Next Micro-Steps (Low Commitment)

1. Read the full Hermes llm-wiki SKILL.md and Karpathy gist side-by-side with the existing Raw to Wiki Compilation and Wiki Health Checks pages.
2. Draft a first version of wiki/SCHEMA.md that captures current practice plus the most useful new conventions (frontmatter fields, provenance syntax, tag taxonomy starter set).
3. Once the synthesis-workbench skill is correctly aligned with the root workbench/ structure, use it to create the next few topics and observe what additional automation would feel natural.
4. Log any recurring pain points during the next 5–10 L3→L2→L1 cycles; those become the spec for the first wiki-oriented skills.
5. Revisit this entry after the next major synthesis (e.g., finishing the Claude Cowork ambitious blueprint work) and decide whether Path 1 + skills or Path 2 feels like the right increment.

Open Questions

• Where should the “long tail” live — inside the main wiki with lighter standards, in a parallel compiler layer, or primarily in workbench topic indexes until they earn promotion?
• How much of the lint/audit/cross-ref work should be skill-driven vs. simple scripts vs. occasional Hermes sessions?
• What does “self-updating” actually mean in practice for this specific voice and Five Dimensions commitment — nightly digests? On-ingest proposals? Weekly contradiction reports?
• How does this interact with the public Quartz surface and the desire for a meta-dashboard / Start Here experience?

This entry exists primarily as a seed. The paths are options to be stress-tested against real usage rather than implemented on principle.

Logged: 2026-05-23 as part of ongoing architecture and skills exploration.
Related: Journal Index · Raw to Wiki Compilation · Wiki Health Checks · Claude Cowork synthesis in workbench.

Journal Calendar · Back to Journal