LLM-Wiki Skills

Portable Agent Skills for coding agents that help users understand, choose, build, migrate, operate, evaluate and govern LLM-Wiki systems.

LLM-Wiki is a pattern where an agent compiles raw sources into a persistent, human-readable, git-versioned Markdown wiki:

raw/  ->  wiki/  ->  AGENTS.md / CLAUDE.md / skills

This repository packages that pattern as installable skills for Claude Code, Codex, Cursor, OpenCode and other Agent Skills-compatible coding agents.

Install

Start with the lane that matches how you will use the wiki:

External starter (recommended)

Create a ready local vault with the supported external-starter profile:

npx llm-wiki-starter

The guided setup asks which agent you use, where your documents are, whether they can leave the machine and whether to enable public export. It creates my-llm-wiki, copies selected documents to inbox/, installs the nine starter skills and runs preflight.

For scripts or a custom destination, use npx llm-wiki-starter init my-llm-wiki --agent codex. Existing files are not replaced; a conflicting external:build package script stops setup. To delegate setup, paste profiles/external-starter/prompt.md into the agent.

New to CLI tools and Git? Follow LLM-Wiki Agent — start here.

After you approve pages under wiki/public/, run npm run external:build in the vault to create a checked public bundle under dist/.

List all available skills:

npx skills add po4yka/llm-wiki-skills --list

Install the full pack for Claude Code:

npx skills add po4yka/llm-wiki-skills --skill '*' -a claude-code

Install selected advisory skills for Claude Code and Codex:

npx skills add po4yka/llm-wiki-skills \
  --skill llm-wiki-orient \
  --skill llm-wiki-faq \
  --skill llm-wiki-choose \
  --skill llm-wiki-zero-to-working-wiki \
  --skill llm-wiki-setup \
  -a claude-code -a codex

Use one skill without installing it:

npx skills use po4yka/llm-wiki-skills --skill llm-wiki-faq --agent claude-code

Zero to working wiki

The fastest product path is:

install core skills -> create raw/wiki vault -> ingest one source -> ask one reusable question -> run wiki-lint -> review git diff

Use docs/quickstart.md for the full 10-minute flow.

To inspect the value before installing anything, open examples/demo-vault/README.md. It shows the first source, expected ingest result, expected query answer and expected lint contract.

Maintenance ritual

After the first run, keep the wiki alive with a small operating loop:

What this pack covers

The repository contains a full lifecycle skill system for LLM-Wiki adoption:

learn -> answer objections -> choose -> diagnose -> plan -> set up -> migrate -> operate -> audit -> publish/archive -> evolve

It includes skills for:

• explaining the pattern and evidence behind it;
• answering serious criticism: consensus smoothing, context poisoning, cognitive debt, drift, token burn, prompt injection, benchmark vacuum and write-only archive risk;
• designing human-first wiki structures that remain findable and readable without agents;
• choosing ready-made versus custom solutions;
• mapping concrete open-source implementations, implementation archetypes, retrieval stacks, ingestion stacks, MCP integrations and eval tooling;
• setting up local-first, repo-docs, Obsidian, MCP/API, retrieval, ingestion or team workflows;
• migrating existing document sets into raw/ and wiki/ structure;
• running triage, ingest, query and lint operations;
• benchmarking pilot value and evaluating whether the wiki is useful;
• auditing provenance, claim anchors, trust, security threat model and model/data policy;
• refreshing stale sources, redacting private content, publishing and archiving;
• designing team operating models, ownership, review queues, SLOs, onboarding and GitOps governance;
• compiling reusable wiki procedures into installable Agent Skills.

Skill groups

Learn and choose

Technology landscape

Diagnose, plan and evaluate

Implement and migrate

Capture and domain workflows

Operate and trust

Skill and memory governance

Repository contents

Validation, distribution, and operational reports

Run deterministic repository checks:

npm run validate

Verify the pack works with the upstream skills CLI as a local source:

npm run smoke:skills

Generate offline refresh reports:

npm run refresh:reports

The smoke test lists available skills, renders llm-wiki-faq with skills use without launching an agent, and installs that skill into a temporary Claude Code project with skills add --copy -y.

The Product E2E workflow installs the starter tarball in a clean environment, then runs a deterministic ingest, query, lint, redaction and export fixture on Linux and macOS for the Claude Code and Codex layouts. It runs on main, weekly and on demand without provider credentials. To run one matrix cell locally:

npm run test:product-e2e -- --agent codex

The refresh reports write dist/source-refresh-report.md and dist/ecosystem-refresh-report.md. They create review queues for stale source-backed claims and ecosystem registry verification; they do not browse the web and do not update truth claims.

Safety principle

Automate bookkeeping, not belief.

The agent may maintain links, frontmatter, indexes, logs, drafts, MOCs, deduplication candidates and lint reports. Human review remains responsible for synthesis, acceptance of ambiguous claims and promotion from draft to reviewed/verified knowledge.

Freshness and evidence policy

Skills that answer current ecosystem questions should use fresh sources. Skills that argue for adoption should distinguish direct LLM-Wiki evidence, adjacent evidence from RAG/memory/context-engineering work, criticism-derived residual risks, and the user's own pilot metrics.