Start the answer for it

What it is

A language model continues text. The first words of an answer shape everything after them: once a reply opens with the refund, it goes on to explain the refund; once it opens with “Thank you for reaching out”, it goes on in that voice. Starting the answer for it means deciding those first words yourself.

There are two forms. The first is an opening line: “Start the reply with exactly this sentence.” Use it when the first sentence carries the message. The second is a skeleton: labels the answer must fill in, such as Status, Done this week and Need from you, in that order. Use it when the answer has a fixed shape.

Developers once did this with a feature called prefill, writing the beginning of the model’s reply in the request itself. In a chat app you cannot type into the reply, so the opening goes in your message, with an instruction to begin with it. For Claude’s newest models it is now the way on the API too, as the last section explains.

When it helps

Use it when the first line is the point: a refund, a decision, bad news. Readers often read only the first sentence, so make sure it is the one you chose.

Use a skeleton when the answer must look the same every time: weekly status updates, handover notes, ticket summaries. A skeleton also makes answers comparable, because this week’s update has the same four labels as last week’s.

Use it to skip preambles. An answer that has to begin with the formula, the verdict or the first heading cannot begin with “Great question”. And use it when a person or a tool will read the first line differently from the rest, such as a formula that will be pasted straight into a cell.

When it does not

Do not write the conclusion before the model has done the work. If you open an analysis with “The annual plan is cheaper”, the model will defend that sentence whether or not the numbers agree. Fix the opening only when you already know what it should say; otherwise ask for the working first.

A skeleton can be too tight. If every update must have a “Need from you” line, the model may invent a need to fill it. Give each slot an allowed empty answer, such as “nothing”.

For output a program will read, a first line is not enough. Use a schema, as in extract to JSON. And a model can still paraphrase your sentence; when exact wording matters, say “exactly” and check it, or paste the sentence in yourself afterward.

Before and after

Three replies whose opening matters, from different kinds of work. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.

Reply to this customer complaint.

"""
[customer email]
"""

Customer email:
"""
[customer email]
"""

Write a reply to the customer above, whose invoice charged them twice. The refund for the second charge was issued today and will reach their card in three to five business days.

Start the reply with exactly this sentence: "You were charged twice, and the second charge has been refunded today."

Then, in under 120 words, explain when the money will arrive and what we have changed so it does not happen again. Format: plain text, two short paragraphs, signed with my name only. It should sound like a person, not a policy. Do not add an apology paragraph. If the email raises anything else, add a [CHECK] note for me at the end.

The first prompt often gets a reply that opens with thanks and an apology and reaches the refund in the third paragraph. The second puts the refund in the first sentence, where an angry customer will read it.

Write a weekly status update for my project.

Write this week's status update for the warehouse scanner rollout, for the operations director, who reads it on a phone between meetings.

My notes:
"""
[notes]
"""

Fill in this structure exactly, and nothing else:

Status: [green, amber or red] because [one reason]
Done this week: [up to three bullets]
Next week: [up to three bullets]
Need from you: [one decision, or "nothing"]

If my notes do not support a status color, write "[CHECK]" instead of choosing one.

A free-form update comes back as several paragraphs in a different order each week. The skeleton returns four labeled lines the director can read in ten seconds, with the status and its reason first.

How do I count orders over $500 in Excel?

In Excel, column C holds order values and column D holds the region for about 2,000 orders. I need the number of orders over $500 in the North region. The count goes straight to a weekly report, so the formula has to work when rows are added.

Begin your answer with the formula alone on the first line, ready to paste into cell F2. Then explain each part of it in one short sentence.

Do not suggest pivot tables or macros. If the formula would behave differently in Google Sheets, say so in one line.

The question gets a short tutorial with the formula somewhere in the middle, often for made-up column letters. The second puts a formula for your columns on the first line, where you can copy it without reading the rest.

The line to add

For an opening sentence:

Begin your reply with exactly this sentence: "{{first sentence}}" Then continue in the same voice.

For a skeleton, list the labels in order and give each an allowed empty answer:

Fill in this structure exactly, and nothing else:

{{Label}}: [{{what goes here}}, or "none"]
{{Label}}: [{{what goes here}}, or "none"]
{{Label}}: [{{what goes here}}, or "none"]

To skip a preamble without writing the opening yourself, name what comes first: “Start with the formula”, “Start with the decision”, “Start with the first heading”.

In ChatGPT, Claude, Gemini and Copilot

In all four chat apps the reply box belongs to the model, so you start the answer from your own message, as in the examples above. That works the same way everywhere.

On Anthropic’s API, the older method has gone. Anthropic’s guidance says that starting with Claude 4.6 models, prefilled responses on the last assistant turn are no longer supported and return an error, while earlier models still accept them (Migrating away from prefilled responses). For a fixed format it points to structured outputs, and for skipping a preamble to a direct instruction such as “Respond directly without preamble.”

Google’s guidance for Gemini describes the same idea as a completion strategy: give the model the start of the answer, or an example and a response prefix, and let it complete the pattern (Prompt design strategies). Its example starts an essay outline for the model so the rest follows the same layout.

In ChatGPT and Copilot, an opening line or a skeleton in your message is the way to do it.

Related

• Ask for a formatName the shape of the whole answer.Pattern
• Extract to JSONWhen a program reads the output, a schema beats a first line.Pattern
• Weekly status update from notesA fixed shape for the update you send every week.Template
• Slide titles that state the pointEvery slide opens with its message.Template