Prompting
Use AI to help specify a scientific task. You check that the resulting prompt preserves the question and makes consequential choices explicit.
Steps
In the workshop, these steps are the exercise. You can also follow them on your own.
- Load one of the three examples in Step 1 · Start (Synthetic scientific query, A 2025-style prompt, or A bare instruction). You can describe a real task of your own instead, with its evidence and the output you need, as long as nothing in it is unpublished or confidential, because the Workbench sends what you type to a third-party model.
- Press Plan it. The model plans the task and lists, in Step 2 · Decision log, the choices it was about to make for you. Review the log. Resolve the forks, correct anything that changes your question, and delete requirements you did not choose. Then press Convert this log → 2026 prompt.
- Copy the prompt from Step 3 · 2026 prompt. Paste it into a fresh conversation in your own AI with your evidence underneath, and run it. Keep the prompt: Verification starts from it.
The Workbench writes the instructions. It usually refers to your evidence (“the supplied table”) rather than copying it into the prompt, so the prompt alone is not self-contained: check what it carries, and paste the evidence with it.
Step 1 · Start
An example replaces the Start box and clears Steps 2 and 3.
Plan it runs the decision-revealing prompt: you see the plan and every choice it was about to make for you before any of it is baked into a prompt.
Step 2 · Decision log
Your edits are authoritative. A fork you leave showing several options becomes an instruction in the prompt to ask you before proceeding — it is never decided behind your back.
Step 3 · 2026 prompt
The 2026-format prompt appears here. Run it in a fresh chat — or send it back round with Re-plan from this to see what it still leaves open.Nothing on this page is a hidden capability. These are the exact system prompts behind the three buttons — copy any of them into your own model and you have the whole workbench without this app.
This page sends what you type to a third-party model through the Vercel AI Gateway — don't paste anything confidential or unpublished. It is limited to 300 generations per 5 minutes per network address (shared by everyone in the room), because the project is unfunded.
The Synthetic scientific query button in Step 1 · Start puts this card in the Start box. Then press Plan it.
Inspect the result
- Does the prompt preserve the actual question and the strength of the claim being assessed?
- Did the model add assumptions, requirements, or an unnecessarily restrictive procedure?
- Does the prompt say what evidence is available and what claims must rest on?
- Would the requested output support the next action you intend to take?
Record
The reviewed task prompt and the first response. Label the prompt clearly so it can be reused unchanged in Verification.
When to use this
You know the work you want done but need help specifying it clearly.