How Do You Get Real Design Options From Claude?
Ask for three approaches, each built around a different priority, and tell Claude to stop before it builds anything. Assigning the priorities is what makes them really different. Saying "stop" matters because Anthropic warns that Claude can start building when you only wanted ideas.
I need to design [thing]. Read [paths]. Constraints: [list]. Give me three approaches that differ in kind: the simplest that could work, one that still holds at [10x current load], and one that ships by [date]. For each, give the idea in two sentences, what it optimizes for, its main downside, what it would change in this codebase, and the point at which it stops working. If two come out as the same design with small changes, replace one. Also say whether not building this yet is a fair option. Then recommend one and stop. Don't write code or change any files until I pick.
- Simplest that works
- Holds at 10x
- Ships by the date
- Don't build it yet
Try it now
- 1In Claude Code, put
/planin front of the prompt so edits are blocked. In the Claude app, send it as is. - 2Test the spread: state in one sentence how each option differs. If you can't, ask for a real alternative.
- 3If two are close, try both. Run
claude --worktree design-ain one terminal andclaude --worktree design-bin another, and spike each. - 4It worked if you can say what you're giving up with the option you picked.
Use it when
Decisions that are expensive to reverse: data models, API shapes, sync or async boundaries, where state lives. It works in the Claude app and in Claude Code. In the app, add the relevant files first, so the options fit your code.
Skip it when
Small, reversible choices where any reasonable design works, such as a helper function. Three options there just slow you down. And if you already know the design you want, say so. Your preference is information Claude can't get elsewhere.
Why it works
Anthropic's Sonnet 5.5 guide says an open-ended request can lead the model to start producing work when you only wanted ideas or a plan, and that you should say so if that's all you want. Its agent teams page describes the other problem: once one theory is explored, later investigation is biased toward it. Asking for options in one go, each with its own priority, makes Claude spread out before it settles. Plan mode backs this up, because edits stay blocked whatever the prompt says. And if a choice stays close, trying both costs little. Anthropic's docs describe /rewind as making a bolder first attempt low-risk, and worktrees let two attempts run without colliding.
Example
For a notifications system, the three might be: send synchronously in the request, which is simplest but ties page speed to the mail provider; an outbox table drained by a worker, which is more moving parts but survives provider outages; or a hosted queue, which is the least code but a new dependency. Suppose you pick the outbox. The trade-off you accepted is now in the conversation when Claude starts building.
Going further
For a big decision, Anthropic's agent teams docs show a prompt that spawns three teammates to look at a design from different angles: UX, technical architecture and a devil's advocate. It's experimental, needs CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, and uses significantly more tokens. Two traps apply to spiking in worktrees. A worktree starts from your default branch, so set worktree.baseRef to "head" if the spike needs unpushed work. It's also a fresh checkout, so list files like .env in .worktreeinclude and install dependencies there.
Common mistakes
- Accepting three options that are one design with different names. Compare what each would actually change in the codebase.
- Leaving off the instruction to stop, then finding option two half built.
- Picking the most sophisticated option by default. Rule of thumb, not a vendor figure: choose the simplest that meets your stated constraints.
- Spiking in your main checkout. Two attempts at once need separate worktrees, or their edits collide.
Sources
- Prompting Claude Sonnet 5.5, Claude Platform Docs
- Orchestrate teams of Claude Code sessions, Claude Code Docs
- Run parallel sessions with worktrees, Claude Code Docs
- Best practices for Claude Code
- Choose a permission mode, Claude Code Docs
Checked 6 November 2026. Products change often, so confirm details with the vendor.