Prompt Recipes

These recipes are written for humans first. Each one tells you the outcome, where it works, what it needs, and how to stop it.

Speak every response during this session

Outcome: An external assistant speaks every reply through Super Voice Mode until you stop the behavior.

Works with: An external assistant connected to the Super Voice Mode Voice toolset.

Copy this prompt:

Use the Super Voice Mode MCP Speak tool with every response during this session until I say stop.

Stop it: Say Stop speaking every response.

Caution

Do not use this recipe with the in-app Assistant. It already manages spoken playback, so an extra Speak call can produce duplicate audio. After the trial, accepted external spoken replies use the shared weekly Pro allowance.

Keep spoken answers concise

Outcome: Shorter replies that are easier to follow by voice.

Works with: The in-app Assistant or an external assistant.

Copy this prompt:

Lead with the conclusion. Keep spoken answers to three short paragraphs unless I ask for more detail.

Stop it: Say Return to your normal response length.

You can also enable the Concise augment for the in-app Assistant.

Draft without submitting

Outcome: The assistant fills a field but leaves the final submission for you to review.

Works with: An agent-capable backend with Computer Use enabled.

Copy this prompt:

Draft the response in the open field, but do not submit it or press Return.

Stop it: No persistent behavior is created. Review or clear the draft.

Turn a Dictionary entry into a prompt launcher

Outcome: A short spoken trigger expands into a longer, reusable instruction.

Works with: Dictation in any editable app.

1. Open Dashboard → Dictionary.
2. Add a distinctive trigger such as insert research checklist.
3. Paste your complete prompt into the replacement field.
4. Dictate only the trigger when you want the exact prompt pasted.

Choose a trigger you would not say in ordinary conversation. Whole-dictation triggers preserve the prompt exactly and skip AI Correction.

Grill Me: interview a plan to clarity

Outcome: The assistant challenges a plan one decision at a time, resolves dependent choices in the right order, and keeps interviewing until you share a complete understanding.

Works with: Any Assistant backend. Use an agent with codebase access when the plan concerns an existing project.

Copy this prompt:

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one by one. For every question, provide your recommended answer and reasoning. Ask one question at a time. If a question can be answered by exploring the codebase, explore the codebase instead of asking me. Format each question as “Question one of X,” where X is a dynamic estimate that may change as the interview progresses. Continue until no unresolved decisions remain, then summarize the agreed decisions, remaining risks, and implementation order.

Stop it: Say Stop the interview and summarize what we have decided so far.

Why one question at a time?

Later decisions often depend on earlier answers. A single-question sequence lets the assistant revise the remaining question count and avoid asking you to choose details that the codebase or a prior decision has already settled.

Ask for a structured decision

Outcome: A complex answer that is easier to read and hear.

Works with: Any Assistant backend.

Copy this prompt:

Give me the conclusion first. Then separate the question, recommended answer, and reasoning. Handle one decision at a time.

Stop it: No persistent behavior is created unless you ask the assistant to remember it.