░██         ░██         ░███     ░███                      ░██░██       ░██
░██         ░██         ░████   ░████                         ░██
░██         ░██         ░██░██ ░██░██    ░██    ░██    ░██ ░██░██    ░██░██
░██         ░██         ░██ ░████ ░██    ░██    ░██    ░██ ░██░██   ░██ ░██
░██         ░██         ░██  ░██  ░██     ░██  ░████  ░██  ░██░███████  ░██
░██         ░██         ░██       ░██      ░██░██ ░██░██   ░██░██   ░██ ░██
░██████████ ░██████████ ░██       ░██       ░███   ░███    ░██░██    ░██░██

github.com/nvk/llm-wiki

LLM-compiled knowledge bases for any AI agent. Parallel multi-agent research, thesis-driven investigation, source ingestion, wiki compilation, querying, and artifact generation. Ships as a Claude Code plugin or a portable AGENTS.md for Codex and others. Obsidian-compatible.

Changelog

v0.2.0 — Nice Cleanup. Lint is the migration — new C11/C12/C13 checks heal misplaced files and legacy layouts via lint --fix. No migrate command needed. Projects simplified — _project.md manifest → plain WHY.md. Focus sessions removed. 9 subcommands → 5. Legacy manifests auto-migrate via lint --fix. Thesis → research mode — /wiki:research --mode thesis "<claim>" replaces /wiki:thesis. Same logic, no duplication. Old command still works as shim. Hub resolution hardened — resolved_path cached in config, symlink recommended for iCloud. Tilde expansion runs at most once. −8% plugin size (3,933 → 3,610 lines) with zero rationale loss.

v0.1.1 — Project-aware lint and compile. /wiki:lint now validates projects (C8 — manifest frontmatter, <!-- DERIVED --> delimiters, project: frontmatter presence and match, Members freshness, slug format) and surfaces migration candidates in existing wikis (C9 — loose binaries as critical violations, sibling binary pairs, version families via prefix clustering, and a fallback that catches topical clusters in any wiki with ≥3 loose outputs and no projects/ folder). --fix regenerates stale _project.md Members sections, backfills missing project: frontmatter, and rebuilds output/_index.md as a projects-aware listing. /wiki:compile regenerates project manifests as a best-effort tail step and steers new outputs with binary siblings into project folders from the start. Completes the v0.1.0 projects architecture.

v0.1.0 — Projects. Group related outputs (playbooks, images, code, data) into project folders under output/projects/<slug>/. Each project has a _project.md manifest with goal, status, and auto-derived member list. Focus into a project with /wiki focus <slug> and subsequent commands inherit it — same as cd-ing into a directory. Multi-project membership via also_in: frontmatter. Lifecycle (active/archived/retracted) is metadata, never file moves. New /wiki:project command with new, list, show, add, focus, unfocus, archive, retract, rename subcommands. /wiki:research and /wiki:ingest now accept --project <slug> and respect focus. The fuzzy router recognizes project intents ("start a new project", "work on X", "what projects do I have"). Backward compatible — loose outputs still work.

v0.0.20 — Bug fixes. /wiki bare command now resolves correctly — the fuzzy router from v0.0.19 is reachable without typing /wiki:wiki.

v0.0.19 — Fuzzy Intent Router. /wiki now understands natural language — type what you want and it routes to the right subcommand. URLs route to ingest, questions route to query, "research X" routes to research, "where was I" routes to resume. Ambiguous input gets a numbered menu instead of guessing.

v0.0.18 — Query Resume Mode. New --resume flag on /wiki:query reloads context after a session break. Reads interrupted research/thesis sessions, recent log entries, wiki stats, and last-updated articles to produce a "where you left off" briefing. Supports combo mode: --resume alone or --resume + a question.

v0.0.17 — Plan Command + Cross-Wiki Context. New /wiki:plan generates wiki-grounded implementation plans (6-stage pipeline: context assembly → interview → gap research → synthesis → plan → save). Supports --format rfc|adr|spec. New --with <wiki> flag on query, output, and plan enables cross-wiki synthesis — use one wiki's craft knowledge when generating from another (e.g., --with article-writing). Merged /wiki:search into /wiki:query --list. /wiki:ingest now supports --new-topic.

v0.0.16 — Full Path Reporting. Agents now report full absolute paths (not relative) and never indent-wrap them, so file links are always clickable in terminals.

v0.0.15 — X.com Fallback Chain. Explicit fallback for X.com/Twitter links: Grok MCP → FxTwitter API → VxTwitter API → direct fetch. Agents no longer waste time debugging login walls. Install ask-grok-mcp for best results.

v0.0.14 — Hub Init Guard. Agents no longer initialize a hub at ~/wiki/ when config points to iCloud/custom path. Fixes bug where an empty ~/wiki/ directory caused research to write to the wrong location.

v0.0.13 — Hub Resolution: ~/wiki First. Resolution now checks ~/wiki/ before reading config. If ~/wiki/_index.md exists, it's used directly — no config read, no iCloud path issues. Config-based custom paths (iCloud, Dropbox) are the fallback, not the default.

v0.0.12 — Configurable Hub Path. Store wiki on iCloud, Dropbox, or any custom location via /wiki config hub-path <path>. Config at ~/.config/llm-wiki/config.json. Default ~/wiki/ unchanged. v0.0.12-fix: Centralized hub resolution into references/hub-resolution.md — fixes iCloud paths where agents misexpanded tildes in com~apple~CloudDocs or broke on spaces in Mobile Documents.

v0.0.11 — Source Retraction. /wiki:retract removes sources and cleans up all downstream references. --recompile rewrites affected articles, --dry-run previews blast radius. New lint rule C4b catches dangling refs.

v0.0.10 — Research Quality. Session registry for crash recovery, standardized agent prompts, credibility scoring (Phase 2b), progress scoring (0-100) with smart termination, and inter-round plan reflection.

Install

Claude Code (native plugin):

claude plugin install wiki@llm-wiki

OpenAI Codex / Any LLM Agent (idea file):

# Copy AGENTS.md into your agent's context or project root
cp AGENTS.md ~/your-project/AGENTS.md

The AGENTS.md file contains the complete wiki protocol as a single portable document — works with any LLM agent that can read/write files and search the web.

Upgrade

Claude Code — if claude plugin update pulls the latest correctly:

claude plugin update wiki@llm-wiki
# Restart Claude Code to apply

If the update command doesn't pick up the new version (stale marketplace cache), sync manually from the repo:

# Clone or pull the latest
git clone https://github.com/nvk/llm-wiki.git  # or: git -C ~/llm-wiki pull

# Sync plugin files to Claude Code's plugin cache
REPO=~/llm-wiki/claude-plugin
DEST=~/.claude/plugins/cache/llm-wiki/wiki
VERSION=$(grep '"version"' "$REPO/.claude-plugin/plugin.json" | grep -o '[0-9.]*')
rm -rf "$DEST"/*
mkdir -p "$DEST/$VERSION"
cp -R "$REPO/.claude-plugin" "$REPO/commands" "$REPO/skills" "$DEST/$VERSION/"

# Restart Claude Code to apply

AGENTS.md — just pull the latest and replace:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/AGENTS.md > ~/your-project/AGENTS.md

Check your installed version: look for the version in /wiki status output or check ~/.claude/plugins/installed_plugins.json.

New to a topic? One command, from anywhere:

/wiki:research "gut microbiome" --new-topic --min-time 1h

Creates a topic wiki, launches parallel agents, and keeps researching for an hour — drilling into subtopics each round finds. Come back to a fully compiled wiki.

Quick Start

/wiki:research "nutrition" --new-topic            # Create wiki + research in one shot
/wiki:research "gut-brain axis" --wiki nutrition   # Add more research to existing wiki
/wiki:research "fasting" --deep --min-time 2h     # 8 agents, keep going for 2 hours
/wiki:research "keto" --retardmax                 # 10 agents, max speed, ingest everything
/wiki:research "What makes long form articles go viral?" --new-topic  # Question → decompose → playbook
/wiki:thesis "fiber reduces neuroinflammation via SCFAs"  # Thesis-driven: evidence for + against → verdict
/wiki:thesis "cold exposure upregulates BDNF" --min-time 1h  # Deep thesis investigation
/wiki:query "How does fiber affect mood?"         # Ask the wiki
/wiki:query "compare keto and mediterranean" --deep  # Deep cross-referenced answer
/wiki:query --resume                              # Where did I leave off?
/wiki add https://example.com/article             # Fuzzy router detects URL → ingest
/wiki what do we know about CRISPR?               # Fuzzy router detects question → query
/wiki:ingest https://example.com/article          # Manually ingest a source
/wiki:ingest --inbox                              # Process files dropped in inbox/
/wiki:compile                                     # Compile any unprocessed sources
/wiki:output report --topic gut-brain             # Generate a report
/wiki:output slides --retardmax                   # Ship a rough slide deck NOW
/wiki:assess /path/to/my-app --wiki nutrition     # Gap analysis: repo vs wiki vs market
/wiki:lint --fix                                  # Clean up inconsistencies

Commands

Command	Description
/wiki <natural language>	Fuzzy intent router — say what you want and it routes to the right subcommand
/wiki	Show wiki status, stats, and list all topic wikis
/wiki init <name>	Create a topic wiki at ~/wiki/topics/<name>/
/wiki init <name> --local	Create a project-local wiki at .wiki/
/wiki:ingest <source>	Ingest a URL, file path, or quoted text
/wiki:ingest --inbox	Process all files in the topic wiki's inbox/
/wiki:compile	Compile new sources into wiki articles
/wiki:compile --full	Recompile everything from scratch
/wiki:query <question>	Q&A against the wiki (standard depth)
/wiki:query <question> --quick	Fast answer from indexes only
/wiki:query <question> --deep	Thorough — reads everything, checks raw + sibling wikis
/wiki:query <terms> --list	Find content by keyword, tag, or category (replaces old /wiki:search)
/wiki:query --resume	Reload context after a session break — recent activity, stats, last-updated articles
/wiki:plan <goal>	Generate wiki-grounded implementation plan (interview → gap research → phased plan)
/wiki:plan <goal> --quick	Plan from wiki content only — skip interview and gap research
/wiki:plan <goal> --format rfc|adr|spec	Output as RFC, ADR, or tech spec instead of roadmap
/wiki:research <topic>	5 parallel agents: academic, technical, applied, news, contrarian
/wiki:research <topic> --new-topic	Create a topic wiki and start researching — works from any directory
/wiki:research <topic> --min-time 1h	Keep researching in rounds until time budget is spent
/wiki:research <topic> --deep	8 agents: adds historical, adjacent, data/stats
/wiki:research <topic> --retardmax	10 agents: skip planning, max speed, ingest aggressively
/wiki:thesis <claim>	Thesis-driven research: evidence for + against → verdict
/wiki:thesis <claim> --min-time 1h	Multi-round thesis investigation with anti-confirmation-bias
/wiki:lint	Run health checks on the wiki
/wiki:lint --fix	Auto-fix structural issues
/wiki:lint --deep	Web-verify facts and suggest improvements
/wiki:output <type>	Generate: summary, report, study-guide, slides, timeline, glossary, comparison
/wiki:output <type> --retardmax	Ship it now — rough but comprehensive, iterate later
/wiki:assess <path>	Assess a repo against wiki research + market. Gap analysis.
/wiki:assess <path> --retardmax	Wide net — adds adjacent fields and failure analysis

All commands accept --wiki <name> to target a specific topic wiki and --local to target the project wiki. Commands that generate content (query, output, plan) also accept --with <wiki> to load supplementary wikis as cross-wiki context — e.g., --with article-writing applies writing craft knowledge when generating output from a domain wiki.

How It Works

Architecture

~/wiki/                                 # Hub — lightweight, no content
├── wikis.json                          # Registry of all topic wikis
├── _index.md                           # Lists topic wikis with stats
├── log.md                              # Global activity log
└── topics/                             # Each topic is an isolated wiki
    ├── nutrition/                      # Example topic wiki
    │   ├── .obsidian/                  # Obsidian vault config
    │   ├── inbox/                      # Drop zone for this topic
    │   ├── raw/                        # Immutable sources
    │   ├── wiki/                       # Compiled articles
    │   │   ├── concepts/
    │   │   ├── topics/
    │   │   └── references/
    │   ├── output/                     # Generated artifacts
    │   ├── _index.md
    │   ├── config.md
    │   └── log.md
    ├── woodworking/                    # Another topic wiki
    └── ...

The hub is just a registry — no content directories, no .obsidian/. All content lives in topic sub-wikis with isolated indexes and articles. Queries stay focused. The multi-wiki peek finds overlap across topics when relevant.

The Flow

1. Research a topic — parallel agents search the web, ingest sources, and compile articles in one command
2. Ingest additional sources — URLs, files, text, tweets (via Grok MCP), or bulk via inbox
3. Compile raw sources into synthesized wiki articles with cross-references and confidence scores
4. Query the wiki — quick (indexes), standard (articles), or deep (everything)
5. Assess a repo against the wiki — gap analysis: what aligns, what's missing, what the market offers
6. Lint for consistency — broken links, missing indexes, orphan articles
7. Output artifacts — summaries, reports, slides — filed back into the wiki

Key Design

• One topic, one wiki — each research area gets its own sub-wiki with isolated indexes. No cross-topic noise.
• Parallel research agents — 5 standard, 8 deep, 10 retardmax. Each agent searches from a different angle.
• _index.md navigation — every directory has an index. Claude reads indexes first, never scans blindly.
• Articles are synthesized, not copied — they explain, contextualize, cross-reference.
• Raw is immutable — once ingested, sources are never modified.
• Multi-wiki aware — queries peek at sibling wiki indexes for overlap.
• Dual-linking — both [[wikilinks]] (Obsidian) and standard markdown links on every cross-reference. Works everywhere.
• Confidence scoring — articles rated high/medium/low based on source quality and corroboration.
• Structural guardian — auto-checks wiki integrity after operations, fixes trivial issues silently.
• Activity log — log.md tracks every operation, append-only, grep-friendly.
• Zero dependencies — runs entirely on Claude Code built-in tools.

Research Modes

Mode	Flag	Agents	Style
Standard	(default)	5	Academic, technical, applied, news, contrarian
Deep	--deep	8	Adds historical, adjacent fields, data/stats
Retardmax	--retardmax	10	Adds rabbit-hole agents. Skip planning, cast widest net, ingest aggressively, compile fast. Lint later.

Smart input detection — /wiki:research auto-detects whether you're passing a topic or a question:

Input	Detected as	Behavior
"nutrition"	Topic	Standard research — explore the field
"What makes articles go viral?"	Question	Decompose into sub-questions → one agent per sub-question → synthesize → generate playbook → suggest theses

Question mode produces a playbook (actionable output artifact) and suggests testable theses derived from the findings.

Modifiers (combine with any mode):

Flag	What it does
--new-topic	Create a topic wiki from the research topic and start immediately. Works from any directory.
--min-time <duration>	Keep running research rounds until the time budget is spent (30m, 1h, 2h, 4h). Each round drills into gaps the previous round found.
--sources <N>	Sources per round (default: 5, retardmax: 15)

# The full combo — new topic, 2 hours of deep research, from anywhere
/wiki:research "CRISPR gene therapy" --new-topic --deep --min-time 2h

Retardmax mode is inspired by Elisha Long's retardmaxxing philosophy — act first, think later. The antidote to analysis paralysis. Works for both /wiki:research and /wiki:output.

Thesis-Driven Research

Unlike open-ended research, /wiki:thesis starts with a specific claim and evaluates it:

/wiki:thesis "intermittent fasting reduces neuroinflammation via glymphatic upregulation"

How it works:

1. Decomposes the thesis into key variables, testable predictions, and falsification criteria
2. Launches parallel agents — but each agent has the thesis as a FILTER. Irrelevant sources get skipped (this prevents bloat)
3. Agents are split: supporting, opposing, mechanistic, meta/review, adjacent — balanced by design
4. Compiles evidence into wiki articles + a thesis file with evidence tables
5. Delivers a verdict: supported / partially supported / contradicted / insufficient evidence / mixed

Anti-confirmation-bias: When using --min-time, Round 2 automatically focuses harder on the WEAKER side of the evidence. If Round 1 found mostly supporting evidence, Round 2 hunts for counter-evidence.

The thesis is the bloat filter. Sources that don't relate to the claim's variables don't get ingested. Higher skip rate = tighter focus.

Linking: Works Everywhere

Every cross-reference in the wiki uses dual-link format:

[[gut-brain-axis|Gut-Brain Axis]] ([Gut-Brain Axis](../concepts/gut-brain-axis.md))

The wiki is not locked into any tool:

• Obsidian reads the [[wikilink]] — graph view, backlinks panel, quick-open
• Claude Code follows the standard (relative/path.md) link
• GitHub/any markdown viewer renders the standard link as clickable
• No viewer at all — plain markdown, readable in any text editor

Obsidian Integration

Each topic wiki has its own .obsidian/ config and can be opened as an independent vault:

open ~/wiki/topics/nutrition/     # Open in Obsidian — focused graph for one topic

The hub (~/wiki/) has no .obsidian/ to avoid nested vault confusion. If you want a cross-topic view, open ~/wiki/ manually and let Obsidian create its own config.

What works out of the box:

• .obsidian/ config created on init with sane defaults
• [[wikilinks]] power the graph view
• aliases in frontmatter enable search by alternate names
• tags in frontmatter are natively read
• inbox/ works as a drop zone in both Obsidian and the CLI

Claude Code is the compiler. Obsidian is an optional viewer.

Query Depths

Depth	Flag	What it does
Quick	--quick	Reads indexes only. Fastest. For simple lookups.
Standard	(default)	Reads relevant articles + full-text search. For most questions.
Deep	--deep	Reads everything, searches raw sources, peeks sibling wikis. For complex questions.
List	--list	Returns ranked article list instead of synthesized answer. Supports --tag and --category filters.

Credits

• Andrej Karpathy — the LLM wiki concept and idea file
• Elisha Long — retardmaxxing philosophy (act first, think later)
• tobi/qmd — recommended local search engine for scaling beyond ~100 articles
• rvk7895/llm-knowledge-bases — prior art in Claude Code wiki plugins

License

MIT License. Copyright (c) 2026 nvk.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files, to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, subject to the following conditions: the above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.