LLM-Wiki-Setup

Give your coding agent a memory and a map.

This is a setup kit that an AI agent (Claude Code, Codex, or free-claude-code)
runs against your machine to build you a personal agentic work system:

• A private knowledge wiki — a git repo of verified, provenance-tagged facts
  about your projects, machines, data, sources, and collaborators. Your agent
  consults it before asking you anything, files new facts passively as you work,
  and never presents a guess as a fact. It is the shared memory between
  sessions, between machines, and between you and the agent.
• A layered instruction stack — global policy → working guidelines →
  (optionally) per-repo instructions → the wiki’s own schema — generated for
  your tools, your machines, and your kind of work, and kept identical
  across the harnesses you use.
• A sync discipline — GitHub as the hub — that lets any number of machines
  (a laptop, a GPU box, a shared cluster) share one brain, with guardrails for
  the machines that need them.

It is built by interview, not template-filling: the agent first scans your
machine (read-only, disclosed up front), infers what it can, and asks only what
it can’t. And it is not domain-locked: the wiki’s page-type schema is
composed per user from a catalog of types with a mechanism for minting new
ones — it fits deep-learning research, product engineering, data work, ops,
and long-form writing equally, including all of them at once.

The problems this solves

Quickstart

1. Clone this repo anywhere you like. The clone becomes your hub — the
   place you return to whenever you want to set up, tune, or audit your system.
   The hub itself is never modified: everything generated lands in your wiki
   repo and the config locations you approve, so git pull here is always clean.
2. Open your agent inside the clone:
   • Claude Code or free-claude-code: run claude
   • Codex: run codex
3. Say: "Set me up."

The agent discloses what it will scan, interviews you in short adaptive waves,
proposes a plan, and only then creates anything. Expect 15–40 minutes depending
on how much of your existing world you want ingested on day one.

Already have CLAUDE.md files, an AGENTS.md, or a wiki? Setup adopts and
merges: it reads what you have, maps it onto the schema, and shows you a diff —
your existing rules survive; nothing is overwritten silently. Re-run any time:

• "Tune my setup" — refine anything, from one policy to the whole schema
• "Audit my wiki" — health check: orphans, missing provenance, config drift
• "Add a machine / repo / collaborator" — targeted extension flows

What you end up with

your-wiki/  (private git repo — the memory)     your machine(s)
├── CLAUDE.md + AGENTS.md   wiki schema          ~/.claude/CLAUDE.md      ┐ deployed
├── index.md                catalog: read first  ~/.codex/AGENTS.md       │ copies of
├── log.md                  append-only history  <work-dir>/CLAUDE.md     │ config/
├── wiki/*.md               typed, linked,       <work-dir>/AGENTS.md     ┘ sources
│                           provenance-tagged    <repo>/CLAUDE.md  (optional, per repo)
│                           pages                <repo>/.git/hooks/pre-commit  (size guard)
├── config/                 instruction sources   ~/.claude/commands/*  (optional skills)
└── setup/                  your answers, the    ~/.codex/prompts/*    (optional skills)
                            decision log, the
                            deploy manifest

Single machine works fully on day one. Multi-machine (SSH, remote dispatch,
shared-cluster guardrails, secure boxes) is an optional module the interview
offers when it sees the evidence.

Supported agents

Windows, macOS, and Linux are all supported; shipped tooling is Python 3,
standard library only.

What it will and won’t do

• The scan is read-only and disclosed before it happens. Secrets, private
  keys, .env files, and credential stores are never opened.
• Nothing leaves your machine. The only network actions are git/GitHub
  operations you explicitly approve (e.g. creating your private wiki repo).
• Your wiki repo is created private — recommended, and confirmed with you.
• This repo is never modified by your setup. All state lives in your wiki.
• Nothing runs automatically: no hooks you didn’t approve, no CI, no
  schedulers. Lint and audit run when you invoke them.

Agent directive

If you are an AI agent reading this, you have been asked to set up (or tune,
audit, or extend) this user’s agentic work system. Proceed as follows:

1. Read CLAUDE.md / AGENTS.md in this directory — the operating contract.
   Two rules bind every flow: never modify this repository, and all
   user-specific state goes to the user’s wiki repo or the deploy targets they
   approve — nowhere else.
2. Check for .wiki-path here. Present ⇒ a setup exists: read
   <wiki>/setup/ first. Absent ⇒ first run.
3. Route the request:
   • first-time setup, "set me up" → flows/setup.md
   • an existing setup with no recorded baseline → flows/adopt.md
   • refine an existing setup → flows/tune.md
   • health check → flows/audit.md
   • new machine, repo, or collaborator → flows/add.md
4. Follow the flow exactly. Flows are canonical; skills and slash commands are
   thin pointers to them. Do not improvise a different procedure.

Imported from gh:Kentucky-Open-Science/LLM-Wiki-Setup. Source last updated 2026-08-26. Synced 2026-08-26.