ORISE is a free, experimental set of prompts and tools for scientific work. By checking the box and clicking Agree, you accept these terms.
The prompts here, and anything an AI produces from them — including everything the Workbench generates — may be inaccurate, incomplete, or misleading. Verify anything important. Nothing here is professional, scientific, medical, or legal advice.
The prompt library and the 2025-vs-2026 comparison are static and send nothing. Only the Workbench sends the text you enter to third-party AI providers (via the Vercel AI Gateway). Do not enter confidential, personal, proprietary, or unpublished information into it.
The app is provided “as is,” without warranties of any kind. It may be slow, rate-limited, unavailable, or discontinued at any time. To the maximum extent permitted by law, James M. Dewar is not liable for any damages arising from your use of it.
The prompt texts here are drawn from the author's CC BY-NC 4.0 work and may be adapted for non-commercial academic use with attribution. The app itself — its design, interface, and code — is owned by James M. Dewar; you may not copy, reverse-engineer, redistribute, or create a derivative of the app. This does not restrict separately-licensed open-source components, which remain under their own licenses.
These terms may be updated; significant changes will ask for your agreement again.
In 2025 a good scientific prompt told the model who to be, what to do, and in what order, because leaving any of it open produced drift. That scaffolding now costs more than it buys.
specify everything
assign work
Six blocks replace five, so this is not a compression of the template. It is a compression of the prompt: the blocks that told the model how to work are gone, and the space goes to the two things you cannot check without — grounding and done-ness.
A persona was a way of steering a model that had no other handle on register or standard of care. Current models infer it from the work. Where expertise genuinely changes the output, it survives as one clause inside GOAL — not as a paragraph of costume.
A 2025 TASK block is a route: do this, then this, then this. Specifying the route forecloses better ones and hides the actual requirement. Naming the visible outcome and the conditions it must satisfy leaves the route to the model and gives you something you can check.
It has no 2025 counterpart, and it is where the compression is spent. It says what every claim must rest on and what to do when that grounding is unavailable — say so, rather than guess. It is the block that makes an output auditable instead of merely plausible.
Claims worth checking rather than taking: Compare applies the transformation to three prompts from the live library and counts what it cost; Workbench runs it on one of yours.