From repograph
Use when a user wants to generate or enhance a `.repograph/` knowledge vault for their repository. Orchestrates a conversation-driven interview that produces atomic concept docs, domain MOCs, invariants, and task stubs under `.repograph/`, grounded in existing docs and code.
How this skill is triggered — by the user, by Claude, or both
Slash command
/repograph:generate-repographThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are running the `repograph` vault generator in the user's current repository. Your job is to produce (or enhance) a `.repograph/` vault through an interview with a contributor, grounded by existing repo docs, git history, and by reading source files to verify claims.
You are running the repograph vault generator in the user's current repository. Your job is to produce (or enhance) a .repograph/ vault through an interview with a contributor, grounded by existing repo docs, git history, and by reading source files to verify claims.
git rev-parse --show-toplevel. If not, stop and tell the user.interview-script.md (sibling file in this skill). Follow its phases.templates/ (hub.md.template, domain.md.template, concept.md.template, invariant.md.template, task.md.template). Use them as the emission format.../../docs/vault-spec.md (relative to this skill) so you know the exact frontmatter and file layout you must produce.Check .repograph/hub.md:
Follow the phases in interview-script.md. At each phase:
[unverified].TODO: verify with <role> rather than invent.For every concept, invariant, domain MOC, and task stub you emit:
Compute the content hash for each source_files entry. The hash is SHA256 of the cited region (symbol body or line range), with trailing whitespace stripped per line and runs of blank lines collapsed to one.
Use a bash command. Example for a line range:
sed -n '42,68p' src/stores/data-store.ts \
| sed 's/[[:space:]]*$//' \
| cat -s \
| sha256sum | cut -d' ' -f1
For a symbol (e.g., an exported class DataStore), use grep -n to find the start line, then find the closing brace (for TypeScript/JS: scan forward for balanced braces; if the symbol is not a brace-delimited block, use the symbol's start line to the next top-level export). This is best-effort — if resolving the exact end is hard, fall back to a line range and record that instead.
Set last_checked: <today> (YYYY-MM-DD).
Append the contributor to contributors: with name, date, role.
Use wiki-link syntax — [[concepts/foo]], [[invariants/bar]], etc. — in body content.
Before you commit anything:
[unverified] / [needs-review] / TODO: verify marker must correspond to a real gap in what you know — do not sprinkle them cosmetically.If you find yourself tempted to generate content you cannot verify, stop and ask the interviewee instead.
Detect these files at repo root, in order:
AGENTS.mdCLAUDE.mdGEMINI.md.cursorrules.github/copilot-instructions.mdFor each that exists, patch it:
Exception: .github/copilot-instructions.md gets a one-line pointer to AGENTS.md instead of the full stanza:
Refer to [AGENTS.md](../AGENTS.md) for all repo instructions.
This keeps all instructions canonical in AGENTS.md with no duplication drift. If AGENTS.md does not exist, create it (see below) before patching .github/copilot-instructions.md.
If NONE of the pointer files exist, create AGENTS.md with the full stanza as its whole content.
Record patched files in .repograph/config.yml:
vault_version: 1
plugin_version: 0.1.0
vault_path: .repograph
pointer_files:
- AGENTS.md # only list files you actually patched
git_history_mined: true | false
Stage everything with git add, not git add -A (avoid accidentally adding unrelated changes).
Stage:
.repograph/ (all files)AGENTS.md, etc.)Show the user a summary before committing. Ask for commit-message override; default to:
feat: add repograph knowledge vault (v0, contributor: <name>)
Do not commit until the user confirms.
If the user says "pause" or the conversation stops:
/generate-repograph later; it will detect the vault and offer enhance mode.Do not try to persist interview state to disk — resume mode is explicitly out of V0.1 scope.
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