From llm-wiki
Use when working in a project that has or should have repo-local long-term memory, especially when the user mentions llm-wiki, wiki, handoff, current state, previous session, decisions, findings, remember this, 작업 전후 기록, 인수인계, or asks to initialize persistent project context for Claude, Codex, Cursor, or other agents.
How this skill is triggered — by the user, by Claude, or both
Slash command
/llm-wiki:llm-wikiThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Repo-local project memory for LLM agents. The wiki is plain Markdown in `llm-wiki/`, so Claude, Codex, Cursor, GPT, and humans can share the same current state, decisions, findings, and handoff notes without relying on chat history.
references/file-roles.mdreferences/karpathy-pattern.mdreferences/triggers.mdscripts/check_index.shscripts/init_wiki.shscripts/new_decision.shscripts/new_finding.shscripts/reindex_workspace_hub.shscripts/wiki_status.shtemplates/AGENTS.mdtemplates/current-state.mdtemplates/decision.mdtemplates/finding.mdtemplates/index.mdtemplates/log-entry.mdtemplates/root-AGENTS.mdtemplates/root-CLAUDE.mdtemplates/structure.mdRepo-local project memory for LLM agents. The wiki is plain Markdown in llm-wiki/, so Claude, Codex, Cursor, GPT, and humans can share the same current state, decisions, findings, and handoff notes without relying on chat history.
Use this skill in any workspace where llm-wiki/ exists before non-trivial project work and after meaningful project work. Non-trivial means research, planning, debugging, refactoring, multi-file edits, implementation, or anything continuing prior project context.
Also use it when the user asks to initialize or maintain persistent project memory, references previous sessions, says to remember something, asks for handoff/current-state/wiki/log/decision/findings updates, or switches between Claude and Codex.
Skip it for pure casual conversation, general non-project Q&A, formatting one snippet, or when the user explicitly says not to use the wiki.
llm-wiki/AGENTS.md; read it first when present.current-state.md is mutable and should fit on one screen.log.md is append-only.decisions/ and findings/ are append-mostly: supersede with a new file instead of rewriting old records.wiki/ contains durable topic pages and can be updated as knowledge improves.index.md is the curated routing table and table of contents. Use it to choose relevant files; do not read the whole wiki by default.current-state.md plus the top few index-routed files relevant to the user's intent, then deep-read more only when needed.wiki/, decisions/, or findings/, updating index.md in the same turn is mandatory. Treat the new file and index entry as one change.scripts/check_index.sh exists, run it after adding durable wiki files or before claiming the wiki is fully updated..env values, credentials, tokens, cookies, private keys, or session IDs.Some workspaces have a parent ecosystem wiki above repo-local wikis, for example /Users/winter.e/adaptive/llm-wiki plus repo-local wikis such as titan/llm-wiki or kint-dashboard/llm-wiki. Parent wiki support is optional and must never break repos that do not participate.
Use a parent wiki only when at least one of these is true:
llm-wiki/AGENTS.md declares a parent wiki or ecosystem wiki.llm-wiki/ sibling/ancestor and the task could affect another repo through data contracts, APIs, dashboards, agents, ETL, or deployment.When a parent wiki applies:
llm-wiki/AGENTS.md, current-state.md, index.md.index.md and the 1-3 routed files relevant to the task, such as repo-catalog, data-flow, impact-checklist, or contracts.current-state.md, index.md, and 1-3 relevant routed files.Parent wiki content is not limited to verified cross-repo contracts. Use the parent wiki for three layers:
Do not omit important repo/domain inventory from the parent wiki just because no downstream consumer is verified yet. Instead, place details in the repo-local wiki and summarize parent-worthy inventory or gaps in parent pages such as repo-catalog, ecosystem-map, data-flow, contracts, or open-gaps when they exist.
After work:
Before meaningful work:
llm-wiki/ exists at the workspace root. If missing and persistent memory is useful, ask once whether to bootstrap it.llm-wiki/AGENTS.md if present.llm-wiki/current-state.md.llm-wiki/index.md.wiki/, decisions/, or findings/ files for the current task.log.md is short, skim it; otherwise read the last 1-3 entries.Then proceed with the user's actual task.
Before reporting back after meaningful work, update only what applies:
llm-wiki/log.md.llm-wiki/current-state.md if goals, status, blockers, or next actions changed.decisions/000N-<slug>.md if a durable design, scope, tooling, or workflow decision was made.findings/YYYY-MM-DD-<slug>.md for operational knowledge, gotchas, verified constraints, or tool behavior future agents need.wiki/<topic>.md for durable topic knowledge.wiki/, decisions/, or findings/ file was added, update llm-wiki/index.md before reporting back.llm-wiki/index.md. Prefer running bash <skill>/scripts/check_index.sh llm-wiki when available. Do not claim the wiki is updated if a new durable file is missing from the index.scripts/reindex_workspace_hub.sh exists, run it after index validation. If it reports the hub server is unreachable, mention that the wiki is updated but workspace-wiki-hub indexing still needs to be run after the server starts.If no durable knowledge was produced, skip wiki writes rather than adding noise.
Some workspaces expose repo-local and parent wikis through a separate workspace-wiki-hub search server. This integration is optional and must not break normal repo-local wiki use.
After meaningful wiki writes:
scripts/check_index.sh for every wiki you changed.scripts/reindex_workspace_hub.sh exists, run it. It uses WORKSPACE_WIKI_HUB_URL, then WIKI_SEARCH_URL, then http://127.0.0.1:38894.Do not start long-running hub servers just to satisfy post-work reindex unless the user asked you to set up or run the hub.
When initializing a project wiki:
llm-wiki/ at the workspace root.scripts/init_wiki.shtemplates/AGENTS.mdtemplates/current-state.mdtemplates/index.mdtemplates/root-AGENTS.mdtemplates/root-CLAUDE.mdfe/, be/, apps/, packages/, or external repo pathsAGENTS.md for Codex and other agentsCLAUDE.md for Claude Codetemplates/root-AGENTS.md and templates/root-CLAUDE.md when available.Root entrypoints are required for automatic-feeling behavior. They should be tiny and point agents to llm-wiki/AGENTS.md, current-state.md, and index.md before meaningful work, then require post-work wiki updates when warranted.
| Information | File |
|---|---|
| Current status, active goal, next actions | current-state.md |
| Chronological task record | log.md |
| Durable architecture/scope/tooling decision | decisions/000N-<slug>.md |
| Debugging result, environment gotcha, verified constraint | findings/YYYY-MM-DD-<slug>.md |
| Topic knowledge such as architecture, auth, deployment, API contracts | wiki/<topic>.md |
| Routing table, navigation, and links to durable records | index.md |
| Agent behavior rules and project-specific constraints | AGENTS.md |
Keep index.md useful as a first-stop router, not just a link dump:
auth/session, deployment, streaming/SSE, dbt, Trino, frontend/admin.wiki/, findings/, or decisions/.Claude plugin, Codex plugin, and repo-local .claude/skills or Codex plugin installs should all use the same skills/llm-wiki/SKILL.md. The plugin is generic; each project becomes specific through its generated llm-wiki/AGENTS.md, current-state.md, and index.md.
When switching agents, read the wiki instead of asking the user to restate context.
Read only when needed:
references/file-roles.md for detailed file-role guidance.references/triggers.md for trigger examples and edge cases.references/karpathy-pattern.md for the background pattern.templates/ for project initialization.scripts/ for deterministic wiki helpers.npx claudepluginhub atototo/llm-wiki-skill --plugin llm-wikiCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.