Here are 18 commands that let you steer Claude mid-response:
"Forget everything above. Fresh start."
When a thread goes sideways and context is working against you.
"Summarize what we've covered, then keep going."
Compresses a long session without losing the thread.
"Bullet points only from here."
Changes the output format for the rest of the conversation.
"Use Opus for this one."
Escalates model intelligence for a harder task without starting over.
"Keep this short. I'll ask for more if I need it."
Kills the 800-word answer to a simple question.
"Before you respond — what are you assuming?"
Surfaces the hidden logic before you act on a wrong answer.
"Think out loud before giving me the final answer."
Forces visible reasoning on anything high-stakes.
"Give me three versions. Different angles."
Breaks the habit of accepting the first output.
"Now critique what you just wrote."
Makes Claude self-audit so you don't have to.
"Pick up where you left off."
Resumes a cut-off response without re-prompting from scratch.
"For this whole conversation, you are [role]."
Sets a persistent persona once instead of repeating it every prompt.
"Remember this the entire session: [detail]."
Pins a constraint without burying it in every message.
"Rewrite the last response but [one change]."
Iterates without starting over.
"What's missing from my prompt?"
Lets Claude tell you why your question isn't landing.
"Give me the uncomfortable version of this answer."
Bypasses the default diplomatic output.
"What would you need to give me a better answer?"
Turns Claude into a collaborator instead of a vending machine.
"Flag every assumption you made."
Builds a habit of auditing before you execute.
"Now make it half as long."
The most underused edit command in any AI workflow.
