Repo LLM Wiki

/repo-llm-wiki

Generates a small, accurate, committable Markdown wiki inside a repo, written at the altitude humans and agents actually use.

Trigger phrases

• /repo-llm-wiki (with subcommands: init, refresh, lint, story)
• "Build a wiki from this repo"
• "Refresh the repo-llm-wiki output"
• "Lint the repo wiki"
• "Update the wiki for my story"

repo source files (Go, TF, README, workflows, scripts, ...)
    ↓  enumerate (cheap shell + reads — main agent)
inventory (cmd dirs, tf resources, scripts, versions)
    ↓  deep analyse in parallel (subagents)
flow narratives + resource summaries + overview prose
    ↓  decide layout + write pages (main agent)
wiki/  (committed, drill-down, agent-readable)
    ↓  lint
a healthy, queryable knowledge base

Subcommands

Skill directory

This SKILL.md and its sibling files (init.md, refresh.md, lint.md, templates/...) live inside the skill directory: ~/.claude/skills/repo-llm-wiki/, ~/.copilot/skills/repo-llm-wiki/, ~/.cursor/skills/repo-llm-wiki/, or the project-local equivalents. In this repository the canonical copy is skills/repo-llm-wiki/. All templates/... paths in this skill are relative to whichever skill directory the loading tool used.

Execution model

This skill is plain Markdown instructions. The "agent" referenced throughout is the host LLM session running the skill (Claude Code, Cursor, or Copilot).

Two-layer: the main session does the cheap deterministic work (shell enumeration, layout decisions, page writing) and fans out to subagents in parallel for the expensive analysis pass — reading every Lambda's entrypoint chain, classifying Terraform resources, grouping scripts, drafting the overview paragraph. Subagent returns are short, structured payloads the main agent assembles into pages. This keeps style consistent (one writer) and protects the main context (subagent context is discarded after each task).

No external models, no orchestration daemon. On hosts without a subagent primitive, the main agent performs the analysis itself inline; output will be the same but slower and noisier.

Dispatch

1. Identify the subcommand from the invocation args (the first word after /repo-llm-wiki).
2. Route — read the sibling file in this skill directory and follow its instructions exactly:
   • init → init.md
   • refresh → refresh.md
   • lint → lint.md
   • story → story.md
3. If the subcommand is missing or unrecognised, print and stop:

   Usage: /repo-llm-wiki <subcommand>
     init                        — scaffold and generate a new wiki
     refresh                     — regenerate wiki from current HEAD
     lint                        — validate wiki health
     story "<description>"       — update wiki from current branch changes
   

4. Do not proceed beyond this dispatch step before reading the target file.

Git policy (canonical)

Never run git commit or git push after any subcommand — init, refresh, lint, story, or any auto-fix — unless the human explicitly asked to commit in this conversation. Subcommand files reference this policy without restating it. The wiki's own AGENTS.md block carries the same rule for downstream agents.