The 4 modes
Mini-claude changes behavior depending on the mode. Same model, same tools, but a different intent. Switch on the fly with /mode <name>, or start on a mode with MINI_CLAUDE_MODE.
auto: does it, no chatter
The full agent. It calls tools, gets the work done, and doesn’t explain itself. For repetitive work, prototyping, when you already know what you want.
explain: codes and tells you why
Like auto, but it adds two sentences of reasoning for every non-trivial decision. Ideal for moving through an unfamiliar codebase, or learning by doing. Like a tech lead showing you between two meetings.
guided: won’t code, makes you find it
guided mode, Mini-claude never writes code. No tools. Its only goal: ask you the right Socratic question so you find the answer yourself.One question at a time, one line max. For studying, mentoring a junior, or forcing yourself to understand instead of copy-pasting. Asking “just give me the answer” won’t work, it politely points you back to the mode.
review: reads and flags what matters
Read-only: it reads your files but writes, edits and runs nothing. It surfaces the 3-4 things worth attention, security, correctness, maintainability, not the nitpicks. One line per finding, with the location (file:line). Senior-reviewer tone, never pedantic.
Recap table
| Mode | Behavior | For |
|---|---|---|
auto | Full agent, executes without explaining | Repetitive work, prototyping |
explain | Codes + 2-sentence reasoning per non-trivial decision | Unfamiliar codebases |
guided | Refuses to code, asks questions | Mentoring, learning |
review | Reads your code, comments on what matters | Review, learning by reading |
See also: Tools · Slash commands.