agent-sync

name	agent-sync
description	>- Use when this capability is needed.

Agent sync (.cursor ↔ .claude)

Purpose

Reconcile this repo’s Cursor config (.cursor/) with Claude Code config (.claude/) so matching files stay byte-for-byte equal when the user wants them to. Default action is copy from source of truth to the other tree (not merge-by-section unless the user asks).

When to use

• User mentions agent-sync, sync .cursor and .claude, mirror skills, or one editor’s copy is wrong.
• After editing a skill, rule, agent, mcp.json, or CLAUDE.md in one tree and the other should match.

Mandatory: source of truth and scope

If the user has not clearly stated:

1. Direction — which tree wins for this run (cursor → claude or claude → cursor), and
2. Scope — a specific skill name, rule file, agent file, mcp.json, or CLAUDE.md, or explicit “sync all mirrored files”,

then stop and ask (use AskQuestion when available, otherwise plain questions). Do not assume .cursor or .claude is authoritative.

Example prompts:

• “Should .cursor → .claude or .claude → .cursor be the copy direction for this sync?”
• “Which single item should we sync first (skill name, mcp.json, CLAUDE.md, or a rules/agents filename)?”

Mirrored paths (this codebase)

Artifact	Cursor path	Claude path
Skill	.cursor/skills/<name>/SKILL.md	.claude/skills/<name>/SKILL.md
Rule	.cursor/rules/*.mdc	.claude/rules/*.mdc
Agent	.cursor/agents/*.md	.claude/agents/*.md
MCP	.cursor/mcp.json	.claude/mcp.json
Claude project instructions	.cursor/CLAUDE.md (optional)	.claude/CLAUDE.md

Skill directory name <name> must match on both sides (e.g. architecture-adr-sync).

If the user wants parity for CLAUDE.md and only one root has the file, create the missing path on the destination using the source file after they confirm direction.

Efficiency — do not read everything

• Do not open every SKILL.md under both skills/ folders to diff the whole set “just in case.”
• Do work one pair at a time: read only the source and destination files needed for the current basename or skill name (e.g. one SKILL.md pair, or one rules file pair).
• For a full sync request: you may list directory entry names (cheap) to find matching basenames, then for each match, read only that pair, apply copy from SoT, then proceed — still never load unrelated skill bodies in bulk before you need them.

Execution steps

1. Read $ARGUMENTS and the latest user message for direction and target (skill name, mcp.json, CLAUDE.md, or a specific rules/agents filename).
2. If either is missing → ask; do not pick a default direction.
3. Resolve paths using the table above; verify the source file exists before copying.
4. Copy source → destination so files match (overwrite destination). Preserve YAML frontmatter and line endings where possible.
5. Summarize in one short paragraph: direction, which paths were updated, and anything skipped (e.g. missing counterpart).

Repo note (git)

In this project, .claude and .cursor/ are gitignored. Syncing still keeps local Cursor vs Claude Code trees aligned; it does not commit to git unless the user changes ignore rules or copies content into tracked docs.

User input

$ARGUMENTS

Honor explicit direction and file/skill names from arguments first; otherwise fall back to asking.

---

Source: Dmitri-IMT/dmitri — distributed by TomeVault.