Ambiguity resolver prompt
The single most common LLM failure mode is making a wrong assumption on a vague request and confidently running with it. This tool appends an ambiguity-detection block to your prompt that forces the model to pause, list the assumptions it would have to make, and ask a capped number of clarifying questions before proceeding — so you catch the misunderstanding before it wastes a turn.
How it works
You optionally paste your existing prompt, then set two knobs: the maximum number of clarifying questions and a confidence threshold. The generated block tells the model to (1) list its assumptions in your chosen format, (2) self-estimate its confidence, (3) ask up to N questions and stop if confidence is below the threshold, or (4) state assumptions briefly and proceed if it is at or above. It also tells the model never to ask about things it can infer from context. The prompt is assembled entirely in your browser.
Tips and notes
- Raise the threshold for costly tasks. If a wrong assumption means rework or a bad action, set the threshold high (85-95%) so the model errs toward asking.
- Lower it for chat. For casual or exploratory use, a lower threshold keeps the model from interrogating the user.
- Cap the questions. One to three high-leverage questions beats a long questionnaire — the limit keeps the interaction snappy.
- Pair with planning. For multi-step work, combine this with a task-planner block so the model clarifies first, then plans.
Why LLMs guess instead of asking
By default, language models are trained and prompted to be helpful on the first attempt. Responding immediately — even on a vague request — is rewarded during training because it looks productive. The result is a model that confidently generates output based on the most plausible interpretation of an underspecified prompt, without signaling that it made a choice.
This is fine when the stakes are low and the assumption is right. It fails badly when:
- A technical request has two equally plausible implementations and the model picks the wrong one
- A creative brief has unstated constraints that make the output wrong despite being well-written
- An agentic task (sending emails, modifying files, calling APIs) proceeds on a faulty premise and produces irreversible results
The ambiguity resolver prompt changes the model’s incentive structure: it explicitly instructs the model to pause, name its assumptions, rate its confidence, and ask before acting when confidence falls below your threshold. This one addition can prevent the most expensive class of LLM error.
What the generated prompt block contains
The builder produces a block that tells the model to:
- Identify ambiguities — list the dimensions along which the request is underspecified
- State assumptions — make the implicit explicit before proceeding
- Estimate confidence — assign a rough percentage reflecting how likely the current interpretation is correct
- Ask or proceed — if confidence is below the threshold, ask the capped number of questions; otherwise state assumptions and act
- Skip inferable gaps — not ask about things that context clearly answers
The output is a plain-text prompt block, no special formatting required, compatible with any LLM that follows system-prompt instructions.
When NOT to use it
For very simple, unambiguous requests where the task has one obvious interpretation, adding an ambiguity block adds latency without benefit. Reserve it for:
- Long or complex tasks where a wrong turn wastes significant effort
- Agentic workflows where actions are hard to undo
- Creative briefs with unstated aesthetic preferences
- Any time you are delegating to an LLM without being available to correct it mid-task