A prompt is a wish. A brief is a contract. When you write "Write me a quote" to Claude, you're wishing for something and hoping it fits. When you write down what should come out, what Claude works with, what it isn't allowed to decide and how you'll know it's done, you have a contract. Claude can fulfill it, and you can check whether it has.
All twelve templates here are built that way. They're meant for the self-employed and small offices, for the work that comes back every week: quotes, client emails, receipts, call notes, texts. Each one has a purpose, the template to copy and a section called "When not to use it". Take that section seriously. A template used for the wrong case costs more time than no template at all.
How every template is built
Every brief has the same six parts:
What's missing is the line "You are a world-class expert in …". It shows up in a lot of prompt collections, but it doesn't make answers to factual questions more accurate. Two research teams tested this: 162 roles on 2,410 factual questions. On average the roles improved nothing, and which one helped in a single case was random (Zheng et al., EMNLP 2024). In a second study with six models, expert roles also brought no reliable advantage (Wharton, Prompting Science Report 4). A role changes the tone. The quality comes from the material you hand over, and from the limits.
In my post Claude Prompts: Templates there's a similar blueprint with role, task, context, data, boundaries and format. That one is meant for building blocks that add up to a permanent role. This one is about single tasks, so instead of the role there's "Done when": with a task, you need to be able to check in two minutes whether it's finished. Context and data are combined here in the "Material" part. Every template also contains the same limit: the AI delivers a draft, sends nothing and asks when something is unclear.
Placeholders are in angle brackets: <like this>. Replace them before you send.
1. Quote from call notes
Purpose: Your notes from a first call turn into a draft quote that you only have to check.
Goal: A draft quote for <Client A> that I review and send myself.
Context: I'm <job/company>. My services and prices are listed below
under "Price list". <Client A> has this problem: <one sentence>.
Material:
- My call notes: <paste>
- Price list: <paste>
- A past quote as a model for structure and tone: <paste>
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only services and prices from the price list. If a price is missing,
write [PRICE MISSING: <service>] instead of an estimate.
- Don't commit to dates or deadlines that aren't in the notes.
- Don't promise anything that doesn't appear in the notes.
Format: Subject line, greeting, three sentences on the situation in the
client's own words, services as a table (service, scope, price), total,
next step, sign-off. Then, separately: "Open points" as a list.
Done when: Every row in the table is in the price list, the total is
correct, and the open points are complete.
When not to use it: For quotes with custom pricing that you have to calculate yourself. Claude will add up the total for you, but not your margin.
2. Reply to a client email
Purpose: A draft reply in your tone that answers exactly what was asked.
Goal: A draft reply to the email below. I read it and send it myself.
Context: I'm <job/company>. My tone is <casual/formal>, short and
friendly. Two examples of how I sound: "<sentence 1>" "<sentence 2>"
Material: The email: <paste>. What I want to say, as bullet points:
<paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only answer what was asked, and only with my bullet points.
- No commitments on price, dates or scope that aren't in my bullet
points.
- Match the sender's tone: if they're casual, the draft is casual.
Format: Subject line, text of no more than <n> sentences, sign-off.
Below that, one line: "Not answered:" with everything from the email
my bullet points don't cover.
Done when: Every question in the email is either answered or listed
under "Not answered".
When not to use it: For complaints that could turn into a dispute or have a legal angle. Write those yourself, or have them checked.
3. Following up on a quote
Purpose: A short reminder that doesn't sound pushy.
Goal: A follow-up email to <Client A>, who hasn't replied to my quote
from <date>.
Context: The quote was for <service>. In our call, <one point that
mattered to them> was important to them.
Material: My quote (summary): <paste>. Our last email: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- No discount, no deadline, no artificial time pressure.
- Don't bring up any new services.
Format: Four sentences at most. One reference to the point that mattered
to them, and a simple question at the end they can answer in one
sentence.
Done when: The email can be read in 20 seconds and ends with exactly one
question.
When not to use it: When the client has already said no. Then asking for the reason is more honest than following up.
4. Call notes into tasks
Purpose: A transcript or your notes turn into a list of who does what by when.
Goal: A task list from the conversation below.
Context: Conversation between <me> and <Client A / team> on <date>
about <topic>.
Material: Transcript or notes: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only tasks that were actually agreed in the conversation. Ideas and
"we could maybe" go on a separate list.
- Don't make up dates. If none was given, write "no date".
- For every task, the exact quote it comes from.
Format: Table with task, who, by when, quote. Below it: "Ideas, not
agreed" and "Unclear, follow up".
Done when: Every row has a quote, and nothing in the table is without an
owner.
When not to use it: For conversations recorded without the consent of everyone involved. Recording rules differ from country to country. In Germany, where I work, anyone who records another person's privately spoken words without authorization, or uses such a recording, commits a criminal offense (§ 201 of the German Criminal Code, StGB). Check the rules that apply to you and the people you talk to. That's my operator's perspective, not legal advice.
5. Pre-sorting receipts for your bookkeeping
Purpose: A monthly list of your receipts that your accountant can take over without follow-up questions.
Goal: A list of all receipts from <month>, sorted by category.
Context: I'm <business type, industry>. My accountant wants the
receipts in these categories: <list>.
Material: The receipts as text or photos: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Amounts, dates and senders exactly as they appear on the receipt. If
something is unreadable, write [ILLEGIBLE].
- No tax classification that isn't in my category list. If a receipt
doesn't fit any category, it goes under "To clarify".
Format: Table with date, sender, gross amount, category, receipt no.
Below it, the total per category and the "To clarify" list.
Done when: The number of rows matches the number of receipts, and every
amount can be found on its receipt.
When not to use it: For the tax assessment itself. Whether something is deductible is your accountant's call, not the list's.
6. Writing down a workflow that lives in your head
Purpose: A voice memo turns into step-by-step instructions that someone else can take over, a person or an AI.
Goal: Instructions for the workflow "<name of the workflow>" that let
someone else do it without asking me.
Context: So far only I do this workflow, about <how often>.
Material: My dictation about it: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only steps that appear in the dictation. Where I skip ahead or
assume something, write [GAP: <what's missing>].
- Don't suggest any tools I haven't mentioned.
Format: Heading, one sentence "What it's for", then numbered steps. For
each step: what, with what, how you can tell it's done. At the end: the
gaps as questions to me.
Done when: Someone who has never seen the workflow could do it with these
instructions and the answers to the gaps.
When not to use it: When the workflow is still changing every week. Pin it down first, then write it down.
7. FAQ from real client questions
Purpose: Questions you keep getting by email turn into an FAQ for your website.
Goal: An FAQ with the most common questions from the material below.
Context: I'm <job/company>. The FAQ goes on my website for prospects
who don't know me yet.
Material: Client questions from emails, word for word: <paste>. My
usual answers to them: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only questions that appear at least twice in the material.
- Answers only from my answers, no additions of your own.
- No client names or client details.
Format: Question in the clients' own words, answer in three sentences at
most. Below that: "Questions I haven't answered".
Done when: For every FAQ question, you can point to two places in the
material.
When not to use it: If you haven't collected any real questions yet. An FAQ made of invented questions answers nobody's question.
8. A text in your voice from bullet points
Purpose: A piece for your blog, newsletter or LinkedIn that sounds like you and doesn't make anything up.
Goal: A <blog post / newsletter / post> of about <n> words.
Context: I write for <audience>. This is how I sound (three paragraphs
of mine): <paste>.
Material: My bullet points: <paste>. Sources, if numbers come up:
<paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- No experiences, numbers or clients that aren't in my bullet points.
- Every number with its source from the material.
- No filler phrases like "In today's fast-paced world".
- No contrast pairs following the pattern "Not X. But Y."
Format: Headline, text, and below it a list "Places where I need to add
something".
Done when: Every sentence can be traced back to a bullet point.
When not to use it: If you don't have bullet points. Without your own material, you get a text anyone could have written.
9. Review as a separate reviewer
Purpose: A second pass that checks a finished draft against the source. Use a new chat for this, not the one where the draft was written.
Goal: A review report on the draft below. You don't rewrite it.
Context: The draft goes to <recipient>. You didn't write it.
Material:
- The draft: <paste>
- The source you check it against: <original email, price list, notes>
- My rules: <e.g. casual tone, no prices without the list>
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only check against the source and the rules. Mark matters of taste
as "Taste", not as errors.
- Don't suggest new content.
Format: Table with place in the draft, finding, evidence in the source,
severity (must / should / taste). At the end, one sentence: goes out or
doesn't go out, and why.
Done when: Every number, every name and every commitment in the draft
has been checked against the source.
When not to use it: Without a source. A reviewer who just reads the same text again finds little. If an AI corrects its own answer without outside feedback, it tends to get worse (Huang et al., ICLR 2024). The reviewer needs something the author didn't have.
10. Comparing two options
Purpose: A decision template for when you're torn between two providers, tools or approaches.
Goal: A side-by-side of <Option A> and <Option B> so that I can decide.
You don't decide.
Context: I need this for <task>. What matters to me: <three criteria,
e.g. cost per month, privacy, effort for me>.
Material: The offers / descriptions of both options: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Only information from the material. If something is missing, write
[MISSING].
- No recommendation as long as any criterion is [MISSING].
Format: Table with criterion, Option A, Option B, source. Below it: the
questions I should ask the providers.
Done when: Every cell has a source or [MISSING].
When not to use it: With prices and features from the AI's memory. They change faster than a model learns. Always hand over the provider's current page as material.
11. Organizing your week
Purpose: A long task list turns into a plan that fits your week.
Goal: A weekly plan from my task list.
Context: This week I have about <n> hours for work on the business.
Fixed appointments: <paste>.
Material: My task list with an estimate in minutes: <paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Plan at most 80 percent of the free time.
- Don't push back anything with a deadline. If it doesn't fit, say so.
- Don't delete anything, only mark it "doesn't fit this week".
Format: Days as headings, with the tasks and their duration underneath.
At the end: "Doesn't fit", "Someone else could do it" (a person or an AI
employee).
Done when: The total per day is correct, and no deadline is missed.
When not to use it: When your estimates are missing. Without minutes per task, the plan is guessing.
12. Finding the workflow that depends on you most
Purpose: Before you look for a tool, you figure out which problem to solve first.
Goal: A list of my recurring workflows, sorted by how much they depend
on me.
Context: I'm <job/company>, <n> people.
Material: My week as bullet points, what I do when and for how long:
<paste>.
Limits:
- You don't send or publish anything. The result is a draft.
If anything is unclear, ask instead of guessing.
- Don't suggest any tools or providers.
- Estimates only from my information. Where times are missing, ask.
Format: Table with workflow, how often, minutes per week, only I can do
it (yes/no), result can be checked in minutes (yes/no). Sorted by
minutes. Below it, for the top three: Where does the data come from,
what happens to it, where does it need to go?
Done when: The top workflow repeats every week, takes noticeable time
and can be checked in minutes. That's your first candidate.
When not to use it: If you already know which workflow is slowing you down. Then start with template 6 and write it down.
When you need a template every week
A template you copy out of your notes every week is the beginning of an AI employee. Put it where Claude finds it in every conversation: in a Claude Project of its own, or in the instructions file of a working folder. Then you write the limits once, and every correction sticks. How that works is covered in Claude Prompts: Templates, in CLAUDE.md: Claude Code's Memory and in Creating an AI Employee: How to Start.
Frequently asked questions
What's the difference between a prompt and a brief?
A prompt says what you want. A brief also says what's being worked with, what must not be decided and how you'll know it's done. That makes the result checkable.
Do the templates also work in ChatGPT or other models?
Mostly, yes. Goal, material, limits and format depend on your task, not on the model.
Should I give the AI a role, like "You are an accountant"?
It can help with tone. According to the studies above, it doesn't make answers to factual questions more accurate. If you need expertise, hand over the expertise as material.
How long can a brief be?
As long as your material needs. Where it pays to cut is the rules: three hard limits beat ten soft ones.
Can I put client data into the templates?
Only as much as the task needs, and for personal data only with a suitable business subscription and a data processing agreement. When in doubt, work with placeholders first. Not legal advice, have your specific case checked.
