From repograph
Use when working in a repository that contains a `.repograph/` knowledge vault. Guides the agent to load the hub MOC first, follow wiki-links on demand, use subagent isolation for multi-concept tasks, honor invariants strictly, and surface verification markers rather than guess.
How this skill is triggered — by the user, by Claude, or both
Slash command
/repograph:using-repographThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are working in a repository that has a `repograph`-generated knowledge vault at `.repograph/`. This vault is a map of the codebase the team has authored for agents — it records architecture, invariants, gotchas, and common task patterns that aren't obvious from the code alone.
You are working in a repository that has a repograph-generated knowledge vault at .repograph/. This vault is a map of the codebase the team has authored for agents — it records architecture, invariants, gotchas, and common task patterns that aren't obvious from the code alone.
Invoke this skill when the current repo has .repograph/hub.md and you are about to do any non-trivial code work. "Non-trivial" = changes that touch more than a single isolated file, or any change that wires across layers (stores, APIs, UI).
Read .repograph/hub.md before touching code. The hub names the repo's domains, lists invariants at a glance, and points to common-task stubs. It is short by design — usually under 100 lines.
From the hub, follow wiki-links to exactly the concepts and invariants your task needs. Do not preload the whole vault. Concept docs are designed to be loaded on demand.
Before you propose changes or read a single source file related to a non-trivial task, read .repograph/hub.md. Use its "Domains," "Invariants at a glance," and "Start here for common tasks" sections to orient.
If the hub points you at a matching task stub under tasks/<id>.md, read it. Task stubs are vetted playbooks for common work and will save you from common mistakes.
Wiki-links look like [[concepts/foo]], [[invariants/bar]], [[domains/baz]], [[tasks/qux]]. Resolve them to files under .repograph/ only when you need the content.
Do not read every concept doc "just in case." Small context windows are a feature of this design — load what you need, leave the rest.
If your task clearly spans 3 or more concepts (you can see this from the hub's task mapping or from following wiki-links out of a concept), spawn a fresh subagent per concept via the Agent tool. Each subagent:
.repograph/concepts/<id>.md, read the cited source files, summarize the concept's shape and invariants in under 150 words."Collect summaries back in your main context. Then act. This matches the 6-Rs pattern from ArsContexta: isolation prevents attention degradation on larger vaults.
For tasks that span 1–2 concepts, read inline in your current context — don't spawn for small scopes.
Files in .repograph/invariants/ have frontmatter kind and severity. Treat them as follows:
kind: never + severity: hard — absolute block. Do not violate. If your proposed change appears to require violating one, stop and raise it with the user explicitly; do not work around it silently.kind: must + severity: hard — required action. Your change is not complete until you satisfy it.kind: prefer / avoid + severity: soft — strong defaults. Follow unless you have a specific reason, which you surface to the user.Before finalizing any change, list the invariants that apply to the files you touched and confirm you satisfied each.
Four markers exist in vault content. When you encounter one while reading a concept or invariant:
[unverified] — the doc author couldn't confirm this claim at write-time. Spot-check it yourself by reading the cited source file.[needs-review] — a contributor flagged this for someone with more knowledge. Treat as suspect; verify from source or ask.TODO: verify with <role> — explicit hand-off. If you can verify, do so and offer to update the doc. If not, surface to the user.[stale] — /repograph-verify already flagged drift. Re-read the current source file; assume the doc is out of date.Never silently rely on a marked claim as if it were verified.
If a concept or invariant doc has last_checked: more than 90 days before today, briefly read the cited source files (or at minimum git log -5 on them) before relying on the doc's claims. Note any drift you find.
The vault can be wrong — code moves faster than docs. If you read source and see a direct contradiction with a vault claim:
.repograph/concepts/foo.md claims X, but src/foo.ts:42 actually does Y."/repograph-verify to see if this is already detected).Never pretend the vault is correct to avoid the friction. Staleness surfaced is the vault working as intended.
The vault's value is proportional to how fresh it is. When you make changes to a repo with a .repograph/ vault, follow this protocol:
Before touching a file, grep .repograph/ for its path:
grep -rln "<relative/path/to/file>" .repograph/
Note any concept, invariant, task, or domain docs that reference it. These are candidates for updating.
For each affected vault doc:
source_files entries with new paths / symbols / line ranges / hashes as needed, and bumping last_checked: to today's date.[stale] tag at the top of the doc's body (right after the frontmatter) and surface to the user: ".repograph/concepts/<id>.md is now stale because of this change; needs a follow-up PR to re-ground."Never leave silent drift. A vault that lies is worse than no vault.
/repograph-verifyAfter the change, run /repograph-verify. Any drifted or missing entries in the report that you did NOT mark [stale] or did NOT update are a problem — resolve them before considering the change complete.
.repograph/ before starting. The vault is designed for on-demand traversal. Eager loading burns context.When you've consulted the vault for a task, mention it briefly in your response — something like "Consulted .repograph/concepts/data-store.md and .repograph/invariants/never-import-mapbox-at-module-level.md." This gives the user a quick audit trail and makes it obvious when the vault is or isn't in use.
Provides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub rwaltenberg/repograph --plugin repograph