Antigravity CLI (agy) has two slash commands specifically to stop the agent from guessing before writing code: /plan and /grill-me. While they sound similar, they solve two very different phases of software design.
Here is how I tend to use each.
I think of /plan as asking a senior engineer to write a technical design doc:
/plan Refactor the webhook processor to use a dead-letter queue on retry exhaustion
When you run /plan, the agent researches your codebase, maps out dependencies, and produces a formal Implementation Plan artifact. It breaks down:
- Every file it intends to create, modify, or delete
- Proposed code diffs and architectural diagrams
- How it plans to verify and test the changes
So does /plan ask any questions? Yes, it does, but only when it hits a blocker or clear ambiguity. If your request is reasonably clear, /plan will infer patterns from your existing codebase, make sensible engineering choices, and highlight any assumptions in an Open Questions or User Review Required section of the plan. You thenreview the blueprint, request tweaks if needed, and approve it to execute.
On the other hand /grill-me shines when I don’t have all the requirements figured out yet:
/grill-me Add per-user rate limiting to our public API
Instead of rushing to draft a blueprint and filling in the blanks itself, /grill-me puts the agent into an interactive, Socratic Q&A mode.
First, it silently reads your codebase so it doesn’t waste your time asking questions the code already answers. Then, it cross-examines your idea using interactive multiple-choice menus right, typically with a recommended default and a custom write-in option:
- Should rate limits be enforced per API key, per user ID, or per IP?
- When Redis is unreachable, should the limiter fail-open or fail-closed?
- Do we need Retry-After headers on 429 Too Many Requests responses?
Each answer shapes the next question until every edge case, failure mode, or product tradeoff are locked down.
This is great for fuzzy ideas, greenfield features, or complex domain logic where you want the agent to poke holes in your requirements before a single line of spec or code is written.
Quick Decision Guide
If you are thinking… “I know what I want, show me which files you’ll touch and how you’ll test it before writing code.” then use /plan.
If it’s rather “I have a rough idea, but I want to make sure I haven’t missed any edge cases or architectural traps.” then start with /grill-me.
Chain Them Together
You don’t have to choose just one! For high-stakes work, the ultimate workflow in agy can involve a two-step combo:
- Start with /grill-me to stress-test your idea and flush out all the hidden edge cases through a fast, interactive terminal interview.
- Follow up with /plan to turn those battle-tested requirements into a concrete, file-by-file implementation blueprint.
By separating discovery (/grill-me) from scoping (/plan) and execution, you’re likely to spend less time reverting AI decisions and more time shipping code that actually fits your architecture.
Why “Plan Mode” is just scoping. Dont forget discovery! was originally published in Google Cloud – Community on Medium, where people are continuing the conversation by highlighting and responding to this story.
Source Credit: https://medium.com/google-cloud/why-plan-mode-is-just-scoping-dont-forget-discovery-a9633382b0df?source=rss—-e52cf94d98af—4
About the Author
