A coding agent skill for building persistent, interlinked wikis from source documents. Drop files into raw/, say /wiki-ingest — the agent reads them, extracts knowledge, and maintains a structured wiki that compounds over time. Extended for novel authoring with an 18-type schema, temporal filtering, reader progressive disclosure, and a full writing pipeline.

This started as a generic knowledge wiki. It's now being extended into a specialized tool for writing Das Kohärenz-Protokoll — a hard sci-fi novel with a DID-protagonist, 13 alter identities, an AI antagonist, 39 chapters, and layered philosophical frameworks. The extensions work for any structurally complex novel.

The project is organized into four major layers, maintaining a clear separation of concerns to prevent context dilution for LLMs.

The static, general-purpose wiki handling any domain — research papers, book notes, business intelligence, or personal knowledge.

• Workflows: /wiki-ingest-knowledge, /wiki-query

Tracks the evolving states and elements of the story, serving as the metadata hub for the writing pipeline.

• Workflows: /wiki-chapter, /wiki-character, /wiki-worldbuild, /wiki-conflict

Models what the reader knows at each point in the novel.

• Workflows: /wiki-reader-model

Governs the underlying rules, agent instructions, and maintenance of the wiki itself.

• Key Files: contradiction-log.md, validation-protocol.md, and session logs.

raw/                    # Immutable source documents — never modify these
processed/              # Source documents that have been ingested successfully
wiki/                   # Agent-maintained wiki layer
  index.md              # Catalog of all pages — updated on every ingest
  log.md                # Append-only chronological record
  overview.md           # Living synthesis across all sources
  knowledge/            # Base layer: concepts, entities, rules, sources, syntheses, timeline
  narrative/            # Novel layer: chapters, characters, locations, conflicts, themes, arcs, beats, dramatica, manuscripts, outlines, timeline
  reader_state/         # Progressive disclosure: foreshadowing, reader-model
  meta/                 # Contradiction log, protocols, archive, ingest
docs/                   # Reference specs (load only when needed)
  jules/                # Developer specifications and architectural proposals
todo/                   # Phased implementation plan with progress tracking
graph/                  # Auto-generated knowledge graph output
tools/                  # Python tools for linting, parsing, graphing, and queries

Document ingestion is handled by the wiki-ingest-workflow. Instead of naive monolithic passes, files are chunked semantically. For each chunk, the agent generates an extraction plan (wiki-decompose) and iteratively performs passes for the Knowledge, Narrative, and Reader State layers. This ensures deep, token-efficient extraction that comprehensively prepares world elements, characters, and rules for downstream writing workflows without dropping context.

Every wiki page declares requires: (what to load for context) and informs: (what to update after changes) in its YAML frontmatter. Reading uses bounded traversal (max-depth 2) to protect the context window. Writing uses unbounded propagation — when a page changes, the agent follows the full informs: chain until every downstream page is updated or confirmed current.

The novel-writing pipeline orchestrates drafting via the /wiki-chapter command, delegating to sub-workflows to generate distinct page types:

Chapter (spec) → Outline (structure) → Beats (atomic scenes) → Manuscript (prose)

Each stage has its own status lifecycle and dedicated workflow within .claude/commands/.

Pages can specify valid_from and valid_until fields referencing timeline events. An agent working on chapter 15 automatically loads only the character states, world-rules, and conflicts that are true at that story-point.

A parallel tracking system models what the reader knows at each chapter. The terminology_permitted list is a one-way ratchet — each chapter inherits the previous chapter's list and may add but never subtract. Foreshadowing strands are tracked across their full lifecycle (planted → reinforced → resolved). Agents can't use concepts the reader hasn't encountered yet.

When pages disagree, resolution follows a strict hierarchy: rule > source > character > chapter > synthesis. Contradictions are always logged in wiki/meta/contradiction-log.md, never silently resolved.

The wiki provides decoupled instructions for different agents based on their persona and tooling constraints.

Agent workflows are structured as slash commands (e.g., /wiki-ingest, /wiki-lint, /wiki-chapter). They run specific toolchain orchestrations, delegating down into individual markdown instructions (using SKILL.md frontmatter) which execute under a 100-line limit for token efficiency.

The /tools/ directory contains optional standalone Python scripts to analyze and maintain the wiki logic (requires ANTHROPIC_API_KEY where LLMs are used).

• tools/lint_deterministic.py: A fast, programmatic tool to check basic YAML schema, frontmatter validity, directory invariants, and orphan states (no API required).
• tools/lint.py: Semantic validation with LLMs checking logic and contextual links.
• tools/check_staleness.py: Traces the informs: chain to identify files that might be out of sync.
• tools/decompose.py: First pass of ingestion, mapping entities before LLM action.

Two-pass build outputting to graph/graph.html:

1. Deterministic — parses all [[wikilinks]] and requires:/informs: edges.
2. Semantic — agent infers implicit relationships → edges tagged INFERRED with confidence scores. Louvain community detection clusters nodes by topic. A SHA256 cache ensures only changed pages are reprocessed. No server, opens locally in any browser.

The original generic wiki layer has been successfully extended with the novel-author architecture. Implementation is following a phased plan defined in todo/README.md.

Requires: Claude Code, Codex, Gemini CLI, or any agent that reads a config file.

git clone https://github.com/netzkontrast/llm-wiki-agent.git
cd llm-wiki-agent

Open in your agent — no API key or Python setup needed (except for tools):

claude      # reads CLAUDE.md + .claude/commands/
codex       # reads AGENTS.md
opencode    # reads AGENTS.md
gemini      # reads GEMINI.md

• Use Obsidian to browse the wiki — follow [[wikilinks]], check graph view, use Dataview for frontmatter queries.
• Use Obsidian Web Clipper to clip web articles directly to raw/.
• The wiki is a git repo — version history for free, diff prose across draft stages.

MIT License — see LICENSE for details.