autowiki

<img src="docs/banner.png" alt="autowiki — documents in, claims validated, contradictions flagged, knowledge base out" width="100%">

Turn documents into a validated, cross-referenced knowledge base. Ingests markdown, PDFs, code, CSVs, logs. Checks that extracted claims match their sources and don't contradict each other.

The wiki that doesn't trust itself. Documents go in, claims get checked against their sources, contradictions get flagged instead of buried.

Install

# Agent Skills (Claude Code, Codex, Cursor, Hermes, OpenCode, OpenClaw)
npx skills add https://github.com/knefenk/autowiki

# Hermes
hermes skills tap add knefenk/autowiki

# Python (supplementary)
pip install autowiki

Skills

Five focused skills, each loaded independently. An agent loading only what it needs uses 0.3K-1.7K tokens instead of one large skill at 2.2K. This split-skill approach follows the pattern from Superpowers (235K stars).

Each skill tells the agent what to do using its native tools -- no scripts required. The ingest skill is a loop: "for each chunk: extract, search KB, update pages, validate, next chunk." If a chunk yields nothing, the agent tries a different approach instead of digging deeper.

How it works

Document → chunk → extract → validate → cross-reference → write pages

1. Read. Document is split with a strategy picked by type: section

boundaries for prose, head+middle+tail for logs, schema+sample for CSV, function boundaries for code. Raw source saved with sha256 so later changes get flagged.

1. Extract. Entities, concepts, and claims pulled from each chunk.

Source provenance tracked (which file, which section).

1. Validate. Claims checked two ways: against the raw source (did we

get the numbers right?) and against existing wiki pages (does this contradict something already known?).

1. Write. Pages created in entities/, concepts/, claims/ with

frontmatter, wikilinks, and confidence scores.

The wiki is plain markdown files. Every wiki includes _navigate.md at its root -- a short file any agent can read to learn how to find and use information in the KB without loading anything else. This pattern comes from agent-kernel.

Platform support

Use

# Start a wiki
autowiki init --path ~/wiki --domain "semiconductor market research"

Then in your agent:

# Ingest a document
"ingest this PDF into ~/wiki"

# Ask questions
"what does my wiki know about TSMC Q1 revenue?"
"show me low-confidence claims about ASML bookings"

Or with the CLI:

autowiki search "P/E valuation" --path ~/wiki
autowiki validate --path ~/wiki
autowiki lint --path ~/wiki

What the validation catches

• Dropped qualifiers. Source says "up to 95%" -- claim says "95%".
• Metric gaming. Abstract says "99% success" -- body reveals it's

pass@8, not pass@1. Flagged at confidence: medium.

• Contradictions. Page A says X, page B says Y about the same thing.

Both marked contested: true, not silently resolved.

• Source drift. Raw file changed after ingest? Sha256 mismatch flags it.
• Simulated data. A log file that looks real but is generated -- flagged

at confidence: low.

Inspirations

• Karpathy's LLM Wiki -- three-layer architecture, markdown-native, wikilinks
• ENPIRE (NVIDIA, 2026) -- EN-PI-R-E closed loop, git-based coordination
• autoresearch (Karpathy, 2026) -- agent-driven experiment loop, program.md pattern
• Superpowers (235K stars) -- split-skill structure, pure SKILL.md approach
• agent-kernel -- AGENTS.md pattern, state/narrative/raw separation
• Loop Engineering -- designing systems that prompt agents, not prompting them directly
• Deli_AutoResearch -- stall detection, heartbeat watchdog, direction diversity

Contributing

Discussions, issues, and PRs welcome. See CONTRIBUTING.md.

License

MIT