The modes of use, and when to use each

Within whatever tool you use, there are four escalating modes, and using the wrong one for a task is a common source of frustration. Match the mode to the task's scope and risk.

1. Tab / inline autocomplete. Single-line to few-line suggestions as you type. Best for flow-state work: boilerplate, known patterns, obvious completions. No execution, minimal risk. You stay in full control, accepting or rejecting each suggestion.

2. Inline chat / "ask" mode. Q&A about your code — explain this function, how do I do X, propose a change you apply manually. Best for understanding and reasoning before you touch anything: onboarding to an unfamiliar codebase, thinking through an approach. Use "ask" specifically when you want no automatic edits.

3. Agent mode. The model plans, edits multiple files, runs commands and tests, reads the output, and iterates until the task is done. Best for well-scoped features, refactors, migrations, and bug fixes that have a reproducible test. This is where real leverage lives — and where verification matters most, because the AI is now changing many things at once.

4. Async / background agents. You assign a task and the agent runs remotely, returning a pull request. Best for parallelizable, well-specified, lower-risk work you can review later — while you do something else.

The escalation is roughly by autonomy, and the right default is the least autonomous mode that fits the task. Anthropic's own guidance captures it: "if you could describe the diff in one sentence, skip the plan" — don't spin up an autonomous agent for a one-line change; just make the edit or take the autocomplete.

A useful decision rule: small and known → autocomplete or direct edit. Uncertain or multi-file → plan, then agent mode. Parallelizable and low-risk → async agent. And a matching risk rule: the more autonomy you hand over, the more rigorously you verify the result (Module 2 and 3). Reaching for agent mode on a trivial change wastes time and adds review burden; reaching for autocomplete on a sprawling refactor leaves you doing the hard part by hand. Fluency here — knowing which gear you're in and why — is half of using these tools well.