rileydrakedesign/wikiforge

Install

Claude Code:   claude plugin install wikiforge --marketplace rileydrakedesign/wikiforge
Cursor:        /add-plugin wikiforge   (once listed)   or add this repo under Customize > Plugins
Any agent:     npx skills add rileydrakedesign/wikiforge
CLI only:      npx wforge-cli init

The plugin ships three standalone skills that work on any folder of markdown pages: lint-wiki, reindex, and wiki-stats. The CLI scaffolds a full wiki with 8 agents and 22 workflow skills.

What you get

npx wforge-cli init

60 seconds later:

my-wiki/
├── CLAUDE.md                   # Schema — how the LLM operates on the wiki
├── wikiforge.yaml              # Boot config (re-run wforge anytime)
├── .forge/
│   ├── agents/                 # 8 agent personas with capability menus
│   ├── skills/                 # 22 workflows across the wiki lifecycle
│   ├── core/                   # /wforge-help navigation skill
│   └── tools/                  # CLI helpers (search, graph, watch, diff)
├── wiki/                       # LLM-maintained: index, log, entities, concepts, sources, comparisons, overview
├── raw/                        # Your source documents (immutable)
├── outputs/                    # Render targets — slides, reports, exports
└── .obsidian/                  # (optional) pre-configured vault — graph view, attachment folder

Open in your LLM tool (claude, cursor, codex). Start with:

"Ingest this source: https://..."

Iris reads it, files entity/concept pages, updates the index and overview, appends to the log, and cross-references against everything already in the wiki. One source typically touches 10–15 pages. Your knowledge compounds.

Why an LLM Wiki?

Not RAG. Sources are never chunked for retrieval. The LLM reads each source once and integrates the knowledge into a persistent, structured wiki — entity pages, concept pages, summaries, comparisons. Cross-references are pre-built. Contradictions are flagged. The synthesis already reflects everything you've ingested.

You curate sources and ask the questions. The LLM does all the bookkeeping — the part humans give up on when wikis decay.

Three layers

Pattern via Karpathy's LLM Wiki gist. WikiForge takes that abstract pattern and generates a concrete, ready-to-run implementation tailored to your domain.

The three Core Loops

These are what make a WikiForge repo a compounding wiki rather than a notes folder. The agents enforce them; you should rarely notice them happening:

1. 🧭 Index-first navigation — agents read wiki/index.md before searching individual pages. The catalog narrows the candidate set; drill-in comes after.
2. ♻️ Answers compound back into the wiki — novel synthesis from queries gets filed back as new pages via the file-answer skill (comparisons → wiki/comparisons/, analyses → wiki/concepts/). Good answers don't evaporate into chat history.
3. 📜 Operations log in parseable form — every ingest, query, lint, and maintenance action appends to wiki/log.md with ## [ISO-timestamp] operation | title. The regular prefix makes the log greppable: grep "^## \[" wiki/log.md | tail -10.

Agents

WikiForge generates eight specialized agent personas. Agents are personas, not separate runtimes — one LLM session adopts whichever role you invoke. Switching agents is switching modes, not handing off between processes.

Only Iris, Quinn, and Linus are enabled by default. Add more during setup.

Skills

Twenty-two workflows organized across the wiki lifecycle:

Each skill ships as an IDE launcher (e.g. /wforge-ingest-source in Claude Code) and a documented workflow under .forge/skills/{phase}/. Workflows are conditional — skills only get installed when their underlying agent is enabled.

Supervision modes

Set a default via the questionnaire; override per-invocation with --mode.

Any workflow accepts --mode deliberate|batch|autonomous at invocation, overriding the config default. Supervision is a per-conversation choice, not a global setting.

Maturity thresholds (autonomous bootstrap)

Autonomous mode is not fully autonomous from day one — sparse wikis can't guide their own research. Tunable in .forge/config.yaml:

Below these, agents fall back to human-guided behavior. As the wiki grows, the system becomes progressively more autonomous.

How the wiki stays maintained

The hard part of any wiki is maintenance. WikiForge bakes it in:

• Lifecycle frontmatter on every page — status, supersedes, superseded_by, last_verified drive the maintenance workflows. Not optional metadata; part of the page contract.
• Auto-fixing lint — Linus repairs the mechanical issues (broken cross-refs to renamed pages, missing index entries, stub creation, frontmatter normalization) and reports the rest. Run with --report-only to see findings without fixing.
• Decay & verify — pages with last_verified past decay_threshold_days (180/90/60 by wiki scale) become candidates for re-verification against current sources.
• Raw immutability — every lint pass re-hashes raw/ files and reports content_hash mismatches under "Modified Sources". Never auto-fixed; re-ingestion is the only safe path.
• Contradictions flagged with confidence: disputed, never silently resolved. The resolve-contradiction skill annotates, splits, supersedes, or reconciles.
• Deduplication via content_hash matching and title/alias similarity. Duplicates are merged, not deleted — non-canonical pages become historical record.
• Schema co-evolution — wforge-revise-schema scans the log for friction signals (ad-hoc page types, frontmatter drift, principles routinely bent) and proposes a schema diff for your approval. The schema is the contract you and the LLM revise together as you learn what the domain needs.

Use with Obsidian

"Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase." — the LLM Wiki gist

When you pick Obsidian-optimized in outputs, WikiForge scaffolds an opt-in .obsidian/ vault:

• app.json with wikilink format + attachmentFolderPath: raw/assets/ per the gist
• graph.json color-coded by wiki subdirectory (entities, concepts, sources, comparisons)
• Sensible core plugins enabled (graph, backlinks, outgoing-link, tag-pane, page-preview, outline)
• README walkthrough for binding Cmd/Ctrl + Shift + D to "Download attachments for current file"

Open the wiki on one side, the LLM agent on the other. As the agent edits pages, watch them update live; browse via [[wikilinks]] and the graph view. Add Dataview from community plugins to run live tables over frontmatter — e.g. "all entities with confidence: disputed, sorted by source_count".

Supported LLM tools

WikiForge generates the appropriate schema file and launcher stubs for your tool:

Setup

wforge init                                       # Full guided questionnaire (recommended)
wforge init --quick                               # Minimal questions, sensible defaults
wforge init --config path/to/wikiforge.yaml       # Regenerate from saved config
wforge init --preset autonomous-wiki              # Start from a community template
wforge init --no-git                              # Skip initial git commit

Community presets

Lineage

The LLM Wiki pattern descends in spirit from Vannevar Bush's Memex (1945) — a personal, curated knowledge store with associative trails between documents. Bush's vision was closer to this than to what the public web became: private, actively curated, with the connections between documents as valuable as the documents themselves. The part Bush couldn't solve was who does the maintenance. The LLM handles that now.

For a sense of scale, look at Tolkien Gateway — thousands of interlinked pages built and maintained by volunteers over years. WikiForge generates the scaffold to build something with that shape, personally, with one human curating and one LLM doing the cross-references.

Development

git clone https://github.com/rileydrakedesign/wikiforge.git
cd wikiforge
npm install
npm run build       # Compile TypeScript → dist/
npm run dev         # Watch mode
npm test            # Vitest suite
npm run lint        # Type-check, no emit

Project structure

src/
  index.ts            # CLI entry (Commander + 14-step generation pipeline)
  cli/                # Questionnaire, quick-start, from-config flows
  generator/          # Per-domain generators (scaffold, schema, agents, skills, obsidian, …)
  config/             # WikiForgeConfig type + defaults + presets registry
  utils/              # fs, git, Handlebars rendering
templates/            # Handlebars .hbs templates — source of truth for generated content
presets/              # Community preset YAML files
tests/                # Vitest suite — smoke + helper coverage

CI runs lint + build + test on Node 18/20/22 for every push and PR.

Project status

Early release. The CLI is feature-complete and tested. Install with npm:

# one-shot, no install
npx wforge-cli init

# or install globally for repeated use
npm install -g wforge-cli
wforge init

To work on WikiForge itself, clone and build from source:

git clone https://github.com/rileydrakedesign/wikiforge.git
cd wikiforge && npm install && npm run build && npm link

Requirements

Node.js ≥ 18

License

MIT © Riley Drake