Step 4 of 6

Choose how the agent should answer

Say the shape of the answer you want, not just the topic. A question, an explanation, a plan and an execution are four different requests, and the agent picks one when you do not.

What happened

What it costs when you skip it

You ask a small question, get nine paragraphs. You ask for an opinion, get an edit to three files. You ask for a fix, get a plan and no fix. Each mismatch costs a round trip, and some of them cost files you did not mean to change.

The expensive case is not verbosity. It is an agent that reads "this feels cluttered" as an instruction to start deleting things.

Do this

1. Pick the mode before you type. Four modes: answer (one or two sentences), explain (short structured reasoning), plan (steps, no execution), execute (do it and report).
2. Put the mode in the first line. "Answer only, no changes." "Explain the tradeoff, do not edit anything." "Plan first, wait for my go."
3. Default to plan for anything destructive. Deletes, bulk renames, migrations and anything that touches a folder you did not create.
4. Keep a stop-and-ask list in the prompt. Files outside the workspace, deletions, anything older than a month, anything it cannot restore.
5. Change mode explicitly, not implicitly. If you want it to act, say "execute the plan" rather than "ok" or "sounds good".
6. Use one chat per mode. A chat that started as questions and moved to execution keeps the reading of the first mode for too long.

ANSWER: one or two sentences. No file changes.
EXPLAIN: what changes, why, main tradeoff. Do not execute.
PLAN: numbered steps, no execution. Wait for my go.
EXECUTE: do exactly the plan above, then list what changed.

Stop and ask before: deleting, renaming, touching files
outside the workspace, or anything that cannot be undone.

Checklist

• The mode is in the first line of the message.
• Destructive work starts in plan mode, not execute mode.
• The stop-and-ask list is present in the session.
• The reply matches the mode you asked for.
• Switching to execution was said explicitly.

Common mistakes

• Controlling behaviour with tone. "Please be careful" is not a constraint. "Do not write outside this folder" is. Ask for behaviour in rules, not adjectives.
• Reading enthusiasm as permission. "I can fix that for you" is an offer, not an action — unless it already acted. Check the files before replying.
• Asking for brevity and then wanting detail. Pick one, per message. A single message that asks for a short answer with full reasoning gets neither.
• Assuming a question is a request. "How would you approach this?" reads as a question to you and as a brief to some agents. Say "answer only" when you mean it.

What you need for this step

Reply gears is in design as a small tool that sets the mode from the key you press to send. Until it ships, the prefixes above do the same job by hand.