How to use the Graft Code Graph Navigator agent skill

What the graft skill makes the agent do

Without the skill, an agent answering "where is rate limiting handled" typically searches text and reads several files. With the skill, it first queries graft/, a graph of the repo made of small markdown nodes. Each node explains one part in prose and names the exact file:line spans it covers, and a wiring graph records who calls what.

The source describes the goal as fewer tokens than rebuilding understanding from source, and catching edges a plain text search would miss. The skill sets six commands for the agent to choose between, and asks it to pick one that fits the task and act on the answer rather than chaining tools.

• ask: locate and understand code, with optional inline source
• grep: find every occurrence of a symbol or pattern
• skeleton: signatures-only view of one file
• callers: who calls a symbol, what it calls, and transitive impact
• map: an orientation tour of an unfamiliar repo
• build / check: refresh the LLM layer, or check that graft/ is not stale in CI

Install the graft skill from the ZIP

Download the ZIP for the skill from this page and unzip it. The package contains SKILL.md. Place the unzipped graft folder in one of these locations:

• Personal skills: ~/.claude/skills/
• Project skills: the .claude/skills/ folder inside a specific repository
• In the Claude apps: upload the ZIP under Settings → Capabilities → Skills

Install the graft CLI and index your repo

The skill assumes the graft CLI is already installed and available on the command line, and that the repository has an indexed graft/ directory. The source says that directory is created by graft build. Without it, the commands will not help.

The source does not give install instructions for the CLI itself, so check the graft project documentation for that step.

• Confirm graft runs from your terminal before starting the agent
• Make sure the repo has a graft/ directory; if not, run graft build from the repo

How to trigger the skill with natural requests

The skill's description tells the agent to use graft for any task in an indexed repo: understanding how something works, finding where code lives, tracing what calls a symbol or what a change breaks, or scoping an edit. You do not need special syntax. Name the task in plain words and mention the repo's code.

Requests that map well to the commands:

• "How does authentication work in this repo?" leads to ask with source
• "Where is rate limiting handled?" leads to ask, then the agent acts on the hit
• "Rename ParseConfig, what will break?" leads to callers with depth 2 or depth all for a refactor
• "Give me the architecture of this codebase" leads to map
• "Find every call site of this constant" leads to grep

Examples of the commands in use

Each command is a single call in the typical case. The skill says most tasks need one call, and that a weak result from one tool should be followed by a different tool, not a reworded question.

For a question about where code lives, the agent runs ask with --source so the matching code comes back in the result. For a rename, it runs callers before editing so it knows every connected file. For an exhaustive search, it runs grep, which groups hits by enclosing symbol.

• ask: use -n N to cap results (default 8) and --in <path> to narrow to a subtree
• grep: use -i for case-insensitive matching and --fixed for a literal string
• callers: default direction in shows who references a symbol; --direction out shows what it depends on
• skeleton <file>: a compact API view of a single file

What it needs and safety notes

Requirements from the source: the graft CLI installed and on the command line, and an indexed graft/ directory created by graft build. The skill runs scripts. It does not use the network. The source says every command needs no API key.

Because the skill runs the graft CLI on your machine, review the CLI and the ZIP contents before installing, as you would any script-running extension. The source also notes that the index can lag behind uncommitted edits in one specific case: when you grep the markdown under graft/ directly, rather than using the graft commands.

• Runs scripts: yes
• Uses network: no
• Licence: MIT
• Limitations: commands depend on graft/; without it they return nothing useful
• Stale markdown: cards under graft/ are rebuilt at the end of a turn, so use the graft tools rather than grepping those files right after editing

Tips for getting good results

Ask for a specific behaviour or symbol name rather than a vague topic. When you already know the symbol, grep it directly and edit at the reported file:line. Before a rename or a multi-file change, have the agent run callers with --depth all to see sibling and downstream files.

If a grep misses, the source advises loosening the pattern to a bare symbol name and retrying graft grep, not falling back to raw grep -rn. In a monorepo, add --in with the sub-project path to ask, grep or callers.

• Use short symbol names in grep, not full signatures
• Narrow monorepo searches with --in <scope>/
• Treat the skill as a navigator: a single well-chosen call is usually enough