Cursor

Cursor is the teammate that doesn't just suggest the next line — it owns the repo. Multi-file refactors, terminal runs, scheduled Cloud Agents, and skills I wrote once that keep paying off every session.

#How Cursor helps me

Copilot is fast fingers. Claude is long-form thinking. Cursor is where the work actually ships — an IDE built around agents that can read the whole repo, run commands, and come back with a diff instead of a paragraph of advice.

I moved from "VS Code + sidebar chat" to Cursor when I realized most of my time wasn't spent typing individual lines. It was spent wiring features across PHP templates, JSON data files, build scripts, and deployment configs. Cursor's agent model matches that reality.

#Workflow

Typical session on nzian.xyz:

1. Open Agent with a concrete goal ("add a team member page for Cursor, match existing patterns").
2. @-reference the files that define the convention — e.g. blog-content/team/github-copilot.md, lib/team/Team.php, the my-team templates.
3. Let the agent explore, edit multiple files, and run validation (php tools/validate-blog-post.php, linters, etc.).
4. Review the diff. Accept what looks right, steer what doesn't.
5. For recurring jobs (blog drafts), I don't even open the IDE — a Cloud Agent automation reads a skill file and commits drafts on a schedule.

The mental model shifted from "AI autocomplete" to "junior dev with repo access and a shell." I still read every diff, but I'm no longer the one manually hunting every reference.

#Strengths

• Whole-repo context. Agent mode reads across blog-content/, lib/, my-team/, .htaccess, and Caddy config in one pass. No more copy-pasting file contents into a chat window.
• Terminal in the loop. It runs git status, php validators, and npm scripts, then fixes what breaks. Copilot never had this; Claude Code gets close, but Cursor keeps it inside the IDE I already have open.
• Persistent instructions. User rules, project rules, and Skills (like my blog-writer skill) survive between sessions. The agent doesn't forget that posts go in blog-content/posts/ with draft: true until I publish.
• Automations that ship while I sleep. Scheduled Cloud Agents on the dev branch draft blog posts Mon/Wed/Fri. I review, publish manually, done.
• MCP integrations. Browser tools for visual checks, Git helpers, Slack — the agent can reach outside the editor when the task needs it.

#Limits

• Agent overreach. Left unchecked, it edits files you didn't ask for or adds "helpful" abstractions. Keep tasks scoped; review diffs before accepting.
• Context windows still bite. Very large repos or marathon sessions can lose thread. Break work into chunks and @-mention the files that matter.
• Model variance. Different models behave differently — one is cautious, another barrels ahead. I pick the model to match the task (fast iteration vs. careful refactor).
• Not a substitute for architecture calls. I still use Claude when I need a patient planning partner. Cursor executes; it doesn't always know what to build.

#Tips

• Write a Skill for anything you do more than twice — blog posts, release checklists, new page scaffolding. One markdown file beats re-explaining conventions every session.
• Pin rules for stack-specific habits (PHP strict types, no commits unless asked, match existing template patterns). Output quality jumps immediately.
• Use @ references liberally. "Follow the pattern in @github-copilot.md" beats a vague "make it similar."
• For automations, keep the prompt short and point at a skill file. The skill is the source of truth; the cron job is just the trigger.
• Disable or gate terminal access on sensitive client repos. Full shell power is the feature and the risk.