draft

LLM Wiki — Draft (local model)

Optional, private, free drafting. When localModel.enabled is true, this skill uses a local model to produce candidate wiki pages into vault/_proposed/ (never wiki/ directly), so a human reviews them via /claude-wiki-pages:review before anything reaches the wiki.

Local models make more mistakes than cloud models — the _proposed/ staging area plus the review gate is exactly the safety margin that makes them usable.

Author drafts to the house voice — see skills/voice: a one-line plain-language definition first, then the typed content in the engineer register, with no marketing language inside the vault.

When to invoke

• The user explicitly wants local/offline/private drafting.
• localModel.enabled is true and an ingest should go through the draft path.

If localModel.enabled is false (the default), do not use this skill — fall back to the normal /claude-wiki-pages:ingest flow with Claude Code.

Configuration

Under localModel in claude-wiki-pages.json:

{
  "localModel": {
    "enabled": true,
    "provider": "ollama",
    "endpoint": "http://localhost:11434",
    "model": "qwen3-coder:30b",
    "draftTarget": "_proposed",
    "tier": "ingest-extract",
    "offlinePolicy": "prefer-local"
  }
}

Read the effective config with bash scripts/engine.sh config --json.

tier is gated per capability (ADR-0018). Only ingest-extract is unlocked today (with qwen3-coder:30b); tier: "draft" is WIRED but BLOCKED until a model clears its quality gate. offlinePolicy governs the Claude→local fallback: off (default) never falls back, prefer-local routes to the approved local tier when Claude is unreachable, strict fails instead of falling back. The deterministic decision lives in bash scripts/engine.sh route (see src/commands/route/CLAUDE.md); the orchestrator consults it. For true-offline drafting with Claude Code stopped, use bash scripts/offline-draft.sh — it applies the same gates and writes the same _proposed/ drafts from a plain shell.