Test one prompt across Claude models and save the winner to Notion

Build me a prompt bench where anyone on our team, not just engineers, can test a prompt against several Claude models at once before we ship it. It runs on Anthropic and Notion. The value is not only the side by side comparison: the shared library we save into is what makes people come back, so treat the Notion side as first class rather than an afterthought.

The main screen is a prompt editor. I write the prompt on one side, and a variables panel lets me define named variables and fill in test values that get substituted into the prompt before it runs. A model picker sits above the editor, populated from List Models in Anthropic so it always reflects the models our account can actually use rather than a hardcoded list. I select between two and four models. There are also basic settings I can adjust: maximum response length, temperature, and an optional system prompt.

When I press Run, send the same resolved prompt to every selected model with Create Message, in parallel, and render the results side by side in one column per model. For each model also call Count Message Tokens and show the input token count under that column so I can weigh answer quality against cost. Show the response time for each column too. Model responses come back as an array of content blocks rather than a plain string, so concatenate the text blocks in order when rendering each column. If one model fails, show the error inside just that column and leave the other columns intact.

Each column has a Save as winner button. Saving creates a page in our shared Notion prompt library with Create a Page, storing the prompt text, the variable definitions, the winning model, the settings used, and the output that won. Let me add a title, a short note on why it won, and tags before it saves. Keep the created Notion page id on the run so a later revision can update that same page.

A Library tab lists everything already saved, read from the Notion database with Query a Data Source, plus a search box that uses Search by Title so people can find a proven prompt instead of starting from scratch. Show title, model, tags, who saved it, and when it was saved. Clicking a row opens the full saved prompt with its output.

From the library I can open a saved prompt back into the bench editor, with its prompt text, variables, model, and settings all restored. After tweaking and re-running it, a Push revision button updates the same Notion page with Update a Page rather than creating a duplicate, and records the new model or settings if they changed.

Several people use the same bench, so show each person only their own recent runs. Keep a Recent runs panel listing that person's runs from the past week with the prompt title, the models compared, and the timestamp, and let them reopen any run to see the columns exactly as they were. Store runs in the app's own storage keyed by the signed in user. The Notion library is shared with everyone; recent runs stay private per person.

Keep the interface plain enough for non technical users. No API language anywhere in the UI: say model, prompt, run, and saved prompt, not endpoints or requests or tokens per minute. Someone in product or marketing should be able to open this, write a prompt, compare four answers, and save the best one without asking an engineer for help.