Daily workflow

Every primitive is now on the table. The mode picker sets how far the agent may go before you look; the model picker and Auto decide which engine does the reasoning; @-mentions and self-gathering retrieval shape what it sees; rules carry the project’s standing facts; run modes - how a run is allowed to execute - and the sandbox set what it can touch; worktrees, subagents, and Cloud Agents multiply it; the agent CLI takes it headless into scripts and CI. Those are Cursor’s current surfaces, as of September 2026, and they shift between releases - every lesson ahead links the primary documentation for the surfaces it names, starting with the mode-switching shortcuts. None of the five pages ahead introduces a new primitive. The subject is the last gap: between knowing a move and reaching for it without thinking.

That gap is where the day-to-day friction of working with an agent actually lives. You know a read-only posture is right while you explore unfamiliar code, but switching the mode is a step, so you don’t - and one day the agent edits a shared module while you were only trying to understand it. You know Auto is the right default and a frontier model is for the hard calls, but under time pressure you leave whatever was last selected and either overpay or under-think. The primitives were the hard part to learn. Making them automatic is the hard part to keep.

This chapter closes that gap in three passes. First it makes the deliberate moves automatic - the posture, tool, and model habits you stop narrating to yourself. Then it moves the repetition out of your head and into the repo, where the agent reads it without being asked. And it ends with a look back across the whole editor-first arc - from a first reviewed change in the sidebar to an unattended agent in CI - to read what the shape of that progression was teaching.

The five moves of this chapter

• The reflexes - the posture, tool, and model habits you stop narrating to yourself - The reflexes worth wearing in
• Choosing the surface - editor, CLI, parallel worktrees, or a Cloud Agent, picked by the shape of the task - Choosing the surface
• Written once - rules, skills, and commands - the standing facts, packaged procedures, and reusable prompt templates you put in the repo instead of in every prompt - Rules, skills, and commands
• The brief - the four parts a prompt can’t be misread without, and when to throw a run away instead of nudging it - Briefing the agent well
• The look back - the whole arc, where Cursor diverged from its terminal-first siblings, and the job underneath all of it - The look back: the whole job

The five run outward from you. The first is about your own hands: the moves you make before you type, made automatic. The second is about placement - the same work is cheap on one surface and painful on another, so you learn to put it where it runs best. The third moves the repetition out of your head entirely and into the repo. The fourth is what’s left after all that: the single task-specific brief, and the discipline of removing ambiguity before a run rather than after it. The fifth isn’t a technique at all. It’s the one place in the course where you stop working and look at what the whole arc was for.