Same prompt, three tabs
ChatGPT, Claude and Gemini will each write a perfectly decent cold email from the same prompt, and the three will differ in ways that are hard to put a finger on. You end up with three tabs and a hunch that the second one was better. Working out why means copying all three into a doc and reading them side by side, which is the step everyone skips.
dassi works in the browser you already use, signed into the same chat accounts you are, so it reads all three answers straight off the page. You ask which one fits the brief. It tells you, and quotes the lines that decided it.
The outreach email, worked through
Take a two-person agency rewriting the prompt behind its first-touch sales email. The brief sits in a Google Doc: who the prospect is, the one offer, a 120-word ceiling, and a ban on “I hope this finds you well.”
- Run the prompt in each chat tool yourself, so every answer sits in its own tab.
- Open dassi’s side panel and ask: “Read the answers in my ChatGPT, Claude and Gemini tabs. Judge each against the brief in my Google Doc and tell me which one a busy founder is most likely to answer.”
- dassi reads the three tabs and the doc and writes a short verdict on each, with the sentences that worked and the ones that broke the brief (the Gemini draft ran to 170 words, for instance).
- Ask for a revised prompt that keeps the strongest opening and the tightest call to action, then run it again. Two rounds is usually where it stops getting better.
We’d skip asking for projected reply rates. The model will cheerfully produce a number, and the number means very little.
The judge has taste of its own
One catch we haven’t solved is that the model doing the judging has preferences. Run dassi on Claude and it may lean toward whichever answer sounds most like Claude, and GPT may do the same for its own family. How strong the effect is, we don’t know. It shows up often enough that we give dassi specific criteria (“under 120 words, one question, no adjectives in the subject line”) rather than asking which answer is best, and when the choice matters we rerun the comparison on a second model. dassi switches between Claude, GPT, Gemini, Grok, DeepSeek, Kimi, OpenRouter models or a local model, so that costs about a minute.
Everything it reads for the comparison, your brief included, goes to the AI provider you picked for dassi, on your own key or through dassi’s relay if you use included credits. The chat tools keep their own copies under their own terms, as they always did.
Where it goes next
The comparison earns its keep right before you hand a job over for good. Once you know which model writes your outreach best, you can teach dassi the weekly routine (pull new leads from the sheet, draft each email with the winning prompt, hold every draft for your approval before anything is sent) and save it as a workflow that runs on that model. The lead list building page covers the list half of that job. The prompt will need rewriting again in six months, when the models change their minds about adjectives.