Your agent guesses instead of asking

Why it guesses

An agent generates the most probable next tokens given your prompt. When you say build me a login, the most probable continuation is a login, not a question about which kind. Asking is a lower-probability move than answering, so unless something tips the scale, the model plows straight into the most likely build and inherits every assumption it did not check.

Training reinforces this. A helpful assistant answers; a hesitant one that keeps asking scores worse, so the model is nudged toward confident action over clarification. The result feels decisive and is often wrong on exactly the decisions that are expensive to reverse, like the auth model that determines your tables and every protected route after it.

The cost is asymmetric. A ten-second question is cheap; discovering after two hundred lines that you wanted magic links, not passwords, is not. The default trades a small certain saving now for a large uncertain cost later, which is the wrong bet on anything that matters.

The manual fix (free, and complete)

The whole fix is to make asking the default on ambiguous work and to force assumptions into the open before code. Paste this into CLAUDE.md or AGENTS.md. It is complete on its own; nothing here is held back for a paid version.

## Ask first

- When a request is ambiguous or a decision is hard to reverse, ask one or two
  sharp questions before writing any code. Do not guess and build.
- Before you start, list the assumptions you are making in one short block. If
  any assumption would change the design, ask instead of assuming.
- Prefer multiple-choice questions ("magic link, password, or SSO?") over open
  ones, so I can answer in a word.
- It is better to ask a needless question than to build the wrong thing. I would
  rather spend ten seconds than throw away two hundred lines.

1. Define when to ask. Require questions when a request is ambiguous or a decision is hard to reverse, before any code.
2. Surface assumptions. Have the agent list its assumptions up front and ask when any of them would change the design.
3. Prefer multiple choice. Ask for A/B/C questions so you can answer in a word and keep momentum.
4. Set the trade-off explicitly. State that a needless question is cheaper than the wrong build, so the agent stops erring toward action.

The kit lets you set a strictness level, from always ask to mostly autonomous, so the ask-first rule matches how much hand-holding you actually want.

The durable fix (the kit)

An ask-first rule in markdown works until the pull to just build wins, and then the agent quietly goes back to guessing. The kit closes that gap with a clarity-lock hook: on an ambiguous task it intercepts the moment before the first code edit and holds the agent at the question until you have answered, so ask-first is enforced at the exact point where the default would otherwise take over.

kitstarter makes your agent ask before it builds, stay lean, and stop looking AI-made, cross-tool through AGENTS.md, at one impulse-priced payment. The written ask-first rule reaches Codex, Cursor, and Antigravity; the clarity-lock engine and the tutor that shows you which question mattered run in Claude Code only.

The real numbers

This is the failure mode our benchmark measures most directly. In the first clean-room suite, five scripted tasks with the same model in both arms, the vanilla agent asked a clarifying question first on two of five tasks; the kit asked on all five. On the ambiguous tasks that meant it built the thing you actually wanted instead of a confident wrong guess.

And the honest flip side: on tasks that were already unambiguous, asking is wasted motion. On a plain todo app the kit was about twice as slow and roughly 75 percent more expensive for no benefit, because there was nothing to clarify. Ask-first is a good trade only where a wrong guess is costly, which is most of real work but not all of it.

It is in Anthropic's own tracker

The tracker frames guessing as a symptom of a deeper habit: acting before checking. These two reports capture it.

• #40541closed2026-03-29CLAUDE.md scope and context-boundary rules ignored after being corrected five or more times in a single conversation. The reporter's summary: the agent prioritizes completing the task quickly over checking constraints before it acts.
• #2901closed2025-07-03The most-upvoted report of the pattern: Claude Code creates deprecated legacy methods instead of editing in place, over-engineers, and ignores single-source-of-truth rules that are written plainly in CLAUDE.md. Closed as not planned with 20+ thumbs-up still on it.

The reporter of #40541 put it exactly: the agent prioritizes completing the task quickly over checking constraints before acting. Guessing instead of asking is that same instinct one step earlier.