The Cold Email Impasse: Finding the Words That Got a Top Professor to Reply
Why generic prompts produce replies a supervisor deletes, and how viva prep prompts and a cold email template give your question a shape worth answering.
The email has been open for forty minutes. It is a paragraph to a professor whose work is the reason for the question, and it has been rewritten six times. viva prep prompts sound like an odd thing to reach for at that moment, and they help for an unglamorous reason: they impose a shape on a request that is otherwise a pile of admiration and context with no ask in it.
The same problem turns up in a viva, where a good answer is rarely about knowing more and usually about ordering what is already known.
Why a generic prompt produces something nobody sends
Ask an AI for an email to a professor and you get three paragraphs of fluent nothing: a compliment with no content, a request with no scope, and a closing that promises to follow up. A supervisor reads four of those a day and can identify the shape instantly.
What changes the result is not a better model, it is a better brief. A prompt that specifies the ask, the length, the context and what counts as a good reply produces something a person could plausibly have written, which is the only version that gets answered.
The stages a research project actually has
Prompts are only useful once they are organised around the work rather than the tool. The prompt library groups them by stage: finding a topic and mapping the literature, breaking a hard concept into analogies, defending a methodology, writing and editing a thesis, checking citations and cross reading sources, responding to reviewer comments and preparing a viva, learning a skill from scratch, and the cold emails and internship applications that sit alongside all of it.
Reaching for the stage you are in beats scrolling a list of two hundred prompts. Most of the time the honest problem is not that the words are missing, it is that the task has not been broken down.
What makes a cold email get answered
The pattern that works is short and slightly uncomfortable, because it asks for something small: a reference to one specific piece of their work, a question narrow enough to answer in two sentences, and a reason you are asking them rather than anyone else.
Length is the easiest thing to get wrong. A paragraph that fits on one phone screen is read; a page with a CV attached and no question is archived. The test to apply before sending is whether a reply would take the reader thirty seconds, because that is the cost you are asking them to accept.
- One specific reference, not three general compliments.
- One question with a boundary, so the answer has a shape.
- One line on why this person, and a clear end to the message.
Practising the viva out loud
A defence rewards structure more than recall. An answer that states the claim, gives the evidence, names the limitation and points at the next step is easy to follow even when the examiner disagrees with it, and it is far easier to deliver under pressure than an improvised one.
The useful exercise is generating the hostile version of your own questions rather than the friendly one, and answering them in four sentences each, out loud. Reading them silently is not the same test.
Unpublished work in a text box
The awkward part of using an AI assistant mid thesis is what the prompt contains: a research question nobody has published, an examiner's criticism, a supervisor's comments. Those are precisely the drafts where an upload feels least comfortable.
The library is static text, which is why it is worth keeping separate from the chat window: the prompts are copied from the page, and the page sends nothing anywhere. What you paste them into is your decision, made afterwards and deliberately.