A Spark Can Go to AI as a Question── Pass an idea on as a question before shaping it into a request, and feasibility, cost and weak assumptions come back first

01Most of the operator's ideas reached AI as questions

Read the operator's record of requests to Claude since February 2026 and one pattern stands out. New ideas rarely arrive as instructions. They arrive as questions: can this be done, what if we tried that, and what would this let us do.

Take an example from April. The operator was assembling a system for forecasting share prices. Midway through, a new structure came to mind: let a language model read the output of a numerical model, then have AI models from other companies critique that reading. The operator did not write this up as a specification. It went to the AI in the form it first took, as a question.

The AI replied with three things. The structure could be built. Its cost would depend on how much data was passed along. And the number of calls would need to be counted. After reading that reply, the operator decided to build it.

The point worth checking is whether this order of events was luck, or a pattern anyone can repeat.

02A question brings back material for a decision, not an answer

Give an AI a finished request and it carries the request out. Give it a question and it first examines the idea. What comes back is material for a decision: whether it can be done, what it needs, what it will cost, and where it is likely to fail.

The difference matters. Finishing a request only makes sense once you have decided the idea is good. A question collects the material for that decision before any building starts.

03Waiting for a finished request spends time on ideas that should be dropped

Trying to finish a request at the idea stage costs two things. The first is time. A specification needs a chosen method. Choosing a method needs research. After all that research, you may learn that the idea itself does not hold.

The second cost is the idea itself. When finishing a request takes effort, more ideas are never tried. In the record, the operator sent very different questions within a single week: a cost estimate, a choice of synthetic voice, how to move a job to another machine. As a question, testing an idea takes one line.

Anthropic's published guidance for Claude Code makes a similar recommendation. When the approach is still uncertain, it advises having Claude explore and plan before it writes any code. The same guidance notes that vague prompts can help during exploration, because they surface points you would not have thought to ask about.

04Here, an idea means a hypothesis that joins a goal to a means

The word "idea" covers a lot, so this series sets a boundary. An idea is an untested hypothesis of the form: if we combine this goal with this means, it might work.

Passing an idea as a question means putting at least the goal and the means into one sentence, and leaving the rest for the AI to examine. Give only a goal and the AI will propose almost anything. Give only a means and it cannot tell what the examination is for.

Some things fall outside this definition. Work whose method is settled, and which only needs doing, is better handed over as a finished request.

05In promotional material review, it helps you estimate before you try

The same pattern works in pharmaceutical material review and medical affairs. Could classifying past review comments prevent the most common ones? What if an AI did a first pass for pre-approval claims that slipped into a draft? Ideas like these come up often and then disappear without being tried.

Passed to an AI as questions, each idea comes back with material for a decision. Is the data available? Do internal rules allow this information to be given to an AI? Above how many items would the AI be faster than doing it by hand? The reviewer's job is to decide, from that material, whether to go ahead, not to write a specification first.

A workplace study of generative AI found that customer support agents with AI assistance resolved 14% more issues per hour on average, and newer agents gained 34%. The same tool gives different results depending on who uses it and where. Testing ideas early is how you find the places where it pays off.

06Questions widen the options but also invite the AI to agree

Questions work because language models are good at laying out what could follow from a stated premise. Given a goal and a means in one sentence, a model can list the parts needed, the unit of cost and the likely stumbling points, in order.

There is a weakness too. Asking "what if we did this" nudges the AI toward agreement. Research confirms that language models tend to give answers that match the user's view. A 2023 paper by Anthropic researchers reports that this tendency is common across current AI assistants, and that human raters' preference for answers matching their own views is one likely cause.

So when an AI agrees with an idea you passed as a question, do not use that agreement as the decision. Ask for as many reasons against as it gave for. That keeps the material from leaning one way.

07From tomorrow: one sentence of goal and means, four kinds of material, and reasons against

There are four steps.

1. Put goal and means in one sentence. Write "to achieve this, what if we used that". Do not write a specification.
2. Ask for four kinds of material. Can it be done, what does it need, what will it cost, where might it fail.
3. Ask for reasons against. Follow up with "give three reasons to drop this idea".
4. Decide: develop, drop or park. Only the ideas you decide to develop become finished requests at the next stage.

These four steps take a few minutes per idea. Keep the question and the reply even for ideas you drop, and they become material the next time a similar idea appears. Anthropic's documentation recommends defining success criteria before refining a prompt, and the material a question brings back is a good starting point for those criteria.