Comparisons
Confusion queries first. Marketplace bait never. Each guide links back to the three pillars.
CLAUDE.md loads on every turn as project memory, skills (SKILL.md) load on demand when a task matches, subagents run a job in their own isolated context window with their own tools, and hooks run a deterministic script on a lifecycle event whether or not the model agrees. Most Claude Code projects use all four at once. Last updated 2026-10-29 by Rakshit Yadav.
Per Cursor's own documentation, a rule (.cursor/rules/*.mdc) is a standing constraint that loads always or by glob, a skill (.cursor/skills/*/SKILL.md) is a playbook the agent opens on demand, and a command (.cursor/commands/*.md) only runs when you type it — see Cursor's docs on commands linked below for the source. Rules constrain, skills teach, commands are typed. Don't put the same text in two places. Last updated 2026-10-31 by Rakshit Yadav.
If the instruction must fire on every turn or on a specific file path, write a rule. If it's a longer playbook the agent should pull in only when the task matches, write a skill. Most solo founders start with ten rules and end up moving eight of them into skills once the rules file gets too long to read. Last updated 2026-10-26 by Rakshit Yadav.
Cursor's native format is .cursor/rules/*.mdc, Claude Code's is CLAUDE.md, and AGENTS.md is the tool-neutral markdown file several agents read. If you use more than one agent, keep one short source of truth and generate the others from it with the Config Converter instead of hand-copying.
The Cursor forum asks this every week. Skills and commands can look like the same markdown. They are not the same job.
An agent is who. A command is the order of who. Confusing them is how people ask one chat to be the whole company.
A command is a markdown prompt that runs only when you type it, and the model interprets it. A hook is a script wired to an agent event in .cursor/hooks.json, and it runs every time that event fires, whether or not the model remembers. Use commands for judgment, hooks for things that must always happen.
A command in Cursor (.cursor/commands/*.md) is a prompt you type with / and run on demand. A rules file (.cursor/rules/*.mdc) is context the model reads automatically, either always or when a glob matches, without you asking. Commands are something you do; rules are something the model already knows. Last updated 2026-10-26 by Rakshit Yadav.
You can write five skills this weekend. That is a valid setup. Here is what you are signing up for if you keep going to 34/49/35 by hand.
Cursor Background Agents give the built-in agent a remote machine to work on, so a long task does not lock your editor. AgenticKit gives that same agent (local or remote) a crew of 46 role prompts, 61 skill files, and 47 commands checked into your repo. One is compute and isolation; the other is the knowledge the agent reads before it acts. Last updated 2026-10-27 by Rakshit Yadav.
“Best” here means: we run them on this product. Not install counts from a marketplace. Twelve to start, 49 if you want the rest.
The hero is /ship. The discipline is knowing when not to run it. These are the commands that survive a real week.
This list is roles you install. It will not rank for “what is Cursor Agent” against cursor.com. It will help you pick a desk.
Cursor skills (.cursor/skills/<name>/SKILL.md) and Claude Code skills (.claude/skills/<name>/SKILL.md) use the same shape — a short YAML frontmatter block with a name and description, followed by a markdown body the agent loads on demand. The folder, the exact fields a client reads, and what else is sitting alongside the skill in each tool's ecosystem are where they diverge. Last updated 2026-10-30 by Rakshit Yadav.
Claude Code skills can be saved two places: a personal folder at ~/.claude/skills/ that every project on your machine can reach, or a project folder at .claude/skills/ that ships with the repo. Cursor only has the second kind — .cursor/skills/ is project-scoped, and there is no personal or global equivalent yet, which is why Cursor's own community forum has open threads asking for one. For a solo founder juggling several SaaS repos, that's the real-world difference: one tool lets you write a skill once and reuse it everywhere, the other makes you copy the file into every repo. Last updated 2026-10-31 by Rakshit Yadav.
Use a skill when the agent needs to know how to do something; use an MCP server when it needs to reach something live, like your database, Stripe account or issue tracker. A skill is a SKILL.md of instructions loaded on demand. MCP is a running process that exposes tools and data. Most solo-founder setups end up with both, and the skill tells the agent how to use the MCP server.
An App Router rules file should be short and glob-scoped to app/**, enforcing the server/client boundary, Server Action conventions and Suspense usage. Full architecture guidance — streaming patterns, parallel routes, Core Web Vitals tradeoffs — belongs in a skill the agent loads on demand, not in a rule that fires on every turn. Last updated 2026-10-26 by Rakshit Yadav.
Stripe code fails quietly: a webhook that verifies against a re-parsed body, a double-charged customer from a retried event, or a subscription that stays "active" after a card decline. The four rules below are glob-scoped to your billing paths so they load only when an agent touches Stripe code, not on every turn. Last updated 2026-10-27 by Rakshit Yadav.
A plugin is a distribution unit — a zip of skills, commands, agents and hooks your team installs with one command. A skill is a playbook that loads when a task matches its description. An MCP server is a live connection to an external system — a database, an API, a SaaS tool — that the model can query or act on mid-conversation. Most real setups use all three together, not one instead of the others. Last updated 2026-10-29 by Rakshit Yadav.
Most published Cursor rules examples are written for Next.js or Python/FastAPI, so teams on React Native, Go, Flutter, or Svelte end up copying a rule file built for a different framework and stripping out what doesn't apply. Below is one annotated starter .mdc per stack — globs, the always-on vs. scoped call, and the two or three bans each ecosystem actually needs. Last updated 2026-10-29 by Rakshit Yadav.
A subagent (.claude/agents/*.md) is a role: a name, a tool grant, and a system prompt that runs in its own isolated context window when the main session or you delegate to it. A slash command (.claude/commands/*.md) is a prompt you trigger by typing /name, read by whichever session is already running. An agent is who does the work; a command is the order you ask for it in. Last updated 2026-10-30 by Rakshit Yadav.
A "Claude Code SEO agent" or "Claude Code growth agent" isn't a separate product — it's a subagent file at .claude/agents/<name>.md with a scoped tool grant, a model choice, and a system prompt for that one job. AgenticKit ships one for each of those roles already: seo-specialist, growth-strategist, email-marketer for lifecycle, support-engineer for ticket triage, and qa-explorer for exploratory QA. Cursor's equivalent lives at .cursor/agents/<name>.md and is invoked by @-mentioning the role in your current chat instead of delegating to an isolated context window. Last updated 2026-11-01 by Rakshit Yadav.
Cursor rules are files with YAML frontmatter stored at .cursor/rules/*.mdc, each one scoped to a glob (or marked always-apply) so only the matching files trigger it. CLAUDE.md is a single plain-markdown file Claude Code loads in full on every turn, with no frontmatter and no built-in path scoping. If you're standardizing on one editor, the two formats aren't drop-in compatible — migrating well means re-checking every rule's scope by hand, not just copy-pasting the prose.
A hook is a deterministic script wired to a lifecycle event — it runs the same way every time, with or without the model's cooperation, and a PreToolUse hook can block an action outright by exiting with status 2. An MCP server is a running process the model talks to mid-conversation to read or change state in something external — a database, a billing account, a GitHub repo. Neither replaces the other: hooks enforce what happens inside the session, MCP servers reach outside it. Last updated 2026-10-31 by Rakshit Yadav.
A slash command is one markdown file under .claude/commands/ that only exists in the project you wrote it in. A plugin is a packaged, versioned bundle — commands, skills, agents and hooks together — installed with one command and reusable across every repo you touch or hand to a teammate. The question isn't which is more powerful; a plugin's commands behave identically to standalone ones. It's whether this workflow needs to leave the repo it was born in. Last updated 2026-10-30 by Rakshit Yadav.
Both tools use the same idea for commands: a markdown file you trigger by typing /name, read into the current session as a prompt. The difference is in storage path, how a team distributes them at scale, and whether the format converts cleanly between the two. If you're running both tools on the same repo and deciding which to standardize on, Claude Code's command format is the one to make primary — here's why. Last updated 2026-10-30 by Rakshit Yadav.
Claude Code skills (.claude/skills/<name>/SKILL.md) load on demand when a task matches their description. Cursor rules (.cursor/rules/*.mdc, or the legacy .cursorrules) load either always-on or by glob/manual trigger, which makes them a different primitive than a skill even though people compare them head to head. Codex's native skills sit closer to Claude Code's model — on-demand, description-gated — but live in their own folder and frontmatter shape. If you're standardizing one repo across tools, know which of these three you're actually choosing between before you pick a format. Last updated 2026-10-31 by Rakshit Yadav.
Connect Sentry's MCP server once with claude mcp add, and Claude Code's debugger subagent can pull a live issue — stack trace, affected release, event count — straight into its context instead of you pasting it in by hand. This is the exact config and the walkthrough for turning a Sentry alert into a verified fix in one pass. Last updated 2026-10-31 by Rakshit Yadav.
ClaudeMarketplaces.com is a browsable directory of individual community Claude Code skills, listed by GitHub author and repo — browse to affaan-m's everything-claude-code pack and you land on a single skill page like team-builder. AgenticKit is a licensed kit that installs a pre-wired team of 46 agents, 61 skills, and 47 commands in one command, with the role-to-skill wiring already done. One is a catalog you assemble from; the other is an assembled team. Last updated 2026-10-31 by Rakshit Yadav.
"Claude skill" means three different things depending on where you heard it: a Claude.ai consumer feature you toggle on in a chat, a repo-native .claude/skills/<name>/SKILL.md file Claude Code loads on demand, or a Claude Agent SDK skill bundled into a custom application built on the API. They share the SKILL.md format and the same on-demand loading idea, but they live in different products with different install paths and different failure modes — including a repo skill silently losing to a same-named plugin skill. Last updated 2026-11-16 by Rakshit Yadav.
Claude Code doesn't publish one official hard cap on simultaneous subagents in its docs — the number people repeat varies by source and by what counts as 'running' (queued vs. actively executing), so treat any specific figure you see online as unverified unless it links to Anthropic's own docs. What is documented and stable: each subagent's frontmatter can pin its own model, and each subagent run gets an isolated context window that returns only a final report to the main thread. Last updated 2026-11-16 by Rakshit Yadav.
A hook is a script wired to a lifecycle event in .claude/settings.json — it runs unconditionally on every matching event, and a PreToolUse hook can block an action outright. A plugin is a packaged, versioned bundle that can contain hooks (plus commands, skills, and agents) so the whole set installs and updates together across repos. They're not alternatives for the same job: a hook is the enforcement mechanism, a plugin is the distribution mechanism for one or more hooks (and everything else) at once. Last updated 2026-11-16 by Rakshit Yadav.
Claude Code's agent teams spin up several subagents that work in parallel inside a single local session, each in its own context window, coordinated by the main agent that dispatched them. Cursor Background Agents take the opposite shape: one agent per task, each running in its own remote cloud sandbox, kicked off from the editor, Slack, or the web and left to finish on its own. One scales out inside a session you're watching; the other scales out across tasks you're not watching. Last updated 2026-11-17 by Rakshit Yadav.
A skill costs a small, fixed amount of context on every turn (its frontmatter description) plus a larger one-time hit only when it actually fires (its body). A plugin costs the sum of whatever skills, commands, and agents it bundles — it has no separate token cost of its own, but it can quietly multiply your fixed tax by installing more skills than you'd add by hand one at a time. This page is narrower than the full primitives comparison: it's only about where the tokens go, not which primitive to reach for. Last updated 2026-11-17 by Rakshit Yadav.
Both tools speak the same Model Context Protocol, so a server that works in one generally works in the other — but the two editors get you there differently. Cursor wants you to hand-edit a mcp.json file; Claude Code wants you to run claude mcp add from the terminal and manages the JSON for you. The servers themselves aren't the hard part — matching scope (project vs. global) and file location across two tools you run against the same repo is. Last updated 2026-11-17 by Rakshit Yadav.