How to write AI work prompts: templates for SMBs

Your team asks ChatGPT to "write an email" and then rewrites the whole thing. What a good work prompt needs, plus three ready-to-use templates for sales, support and operations.

Why does ChatGPT give you such generic answers?

It is Tuesday and your sales rep needs to follow up on a quote she sent last week. She opens ChatGPT and types "write a follow-up email for a client". What comes back opens with "Dear valued customer", talks about "tailored solutions" and closes by offering a discount nobody approved. She deletes most of it and writes it herself.

Support runs into the same thing. Someone pastes in a customer complaint and types "reply to this". The draft apologizes five times and promises a refund before anyone knows whether one applies.

The team concludes that "AI doesn't work for this". Most of the time the problem is the instruction, not the tool.

Why doesn't the AI understand what you ask for?

Because it knows nothing about your company. Anthropic's guide for Claude puts it this way: think of the model as a brilliant but new employee who lacks context on your norms and workflows. Tell a new hire "follow up with that client" and they will ask which client, why, and what they are allowed to offer.

Prompts are written like Google searches. Five words, no context, hoping the model guesses the rest.
What only your team knows is missing: who the customer is, what was quoted, what can be promised and what cannot.
Nobody says what the result should look like. No length, no tone, no channel. An email and a WhatsApp message are not written the same way.
Everyone reinvents the prompt every time. When something works for one person, it stays in their chat history and nobody else benefits.

What makes a good work prompt?

The official guides from Anthropic, OpenAI and Google agree on the basics. Google's guide for Gemini in Workspace boils it down to four areas: persona, task, context and format. Anthropic's adds two more: examples, and explaining the reason behind each instruction. For a small business, that fits into five pieces:

Role: who the AI is for this task. "You are a sales rep at a restaurant-supply distributor in Bogotá" shapes tone and vocabulary.
Task: what it should do, with a clear verb. "Write a WhatsApp follow-up message", not "help me with this client".
Context: the facts of the case and the reason behind each instruction. Anthropic recommends explaining why: "keep it short because they will read it on their phone" works better than a bare "80 words max".
Format and limits: length, channel, tone, structure and what is off-limits. Google suggests concrete constraints, such as a character limit or the number of options you want.
Example: one or more real messages you like. According to Anthropic, examples are one of the most reliable ways to steer format and tone; when you need consistent results, it recommends three to five.

There is a sixth piece that lives outside the prompt: iteration. OpenAI recommends reviewing the first response and refining it with more context. The first attempt is rarely the final one.

Sales template: following up on a quote

Copy this, replace what is in brackets and paste the quote details at the end:

Role: You are a sales rep at [company], which sells [product or service] to [type of customer] in [city or region].
Task: Write a WhatsApp message following up on quote [number], which we sent to [contact name] at [client company] on [date].
Context: The client asked for [what they asked for]. We quoted [summary and amount]. The quote expires on [date]. Their main concern was [delivery time, price or payment terms].
Format: Five lines at most, because they will read it on their phone. Polite and direct (in Spanish, use "usted"), with no "Dear valued customer". End with one concrete question that invites a reply.
Limits: Do not offer discounts, deadlines or terms that are not in the quote. If a detail is missing, leave it in brackets so I can fill it in.
Example of a message that worked for us: [paste a real message here].

The limits line saves the most rework. Without it, the model may fill the gaps with things your company never approved. The same structure works for drafting a proposal: change the task and paste the scope, deliverables and terms into the context.

Support template: replying to a complaint

Complaints are where a weak prompt shows the most. This template prepares a draft for a person to review, not a reply that goes out on its own:

Role: You are on the customer service team at [company]. Your goal is for the customer to feel heard and to know what happens next.
Task: Write a draft reply to the complaint between the <complaint> and </complaint> tags.
Context: Our policy for this case is [warranty, exchange or return policy]. What we have already done: [actions]. The person who can approve exceptions is [role].
Format: One paragraph acknowledging the specific problem without repeated apologies, and one with the next step and the date we will contact them. 120 words at most.
Limits: If the case calls for a refund, discount or compensation, do not promise it: say you are escalating it to [role]. Do not blame the customer or another team.
Before the draft: tell me in one line what the customer is really asking for and whether any information is missing.

The tags are there for a reason. Anthropic recommends using XML-style tags to separate your instructions from the text you paste in, so the model does not mistake a sentence from the customer for an order from you. And when the material is long, such as a full email thread, both Anthropic and Google advise putting it first and leaving the instruction at the end.

Operations template: meeting notes and procedures

In operations, the most common use is turning messy notes into something the team can follow. For meeting minutes:

Role: You are the operations assistant at [company].
Task: Using the notes or transcript inside the <notes> tags, write the minutes of the [topic] meeting held on [date].
Format: Three sections: decisions made, action items (a table with owner, task and due date) and open items for the next meeting.
Limits: Use only what is in the notes. If an action item has no owner or date, write "not set" instead of guessing.

For a procedure, the step-by-step guide most small companies never write down, change the task: "Using this explanation of how we do [process], draft a procedure with purpose, owner, numbered steps and what to do if something goes wrong. At the end, list the questions a new hire would have after reading it." That final list is the most valuable part: it shows the gaps only your team can fill.

How can your team start using prompt templates this week?

Day 1: collect the tasks. Ask each team for the three tasks they use AI for most, or the ones they still write by hand most often.
Day 2: write the first version. One template per task, with the five pieces. The person who does the task writes it, not the manager.
Day 3: test it on real cases. Three to five different cases, including a hard one. Compare with what a person would have written and adjust.
Day 4: store it where everyone can find it. A shared document works. If you use business accounts, Projects in Claude and ChatGPT and Gems in Gemini let you save instructions and reference files for reuse.
Day 5: give it an owner. Someone who updates it when prices, policies or brand tone change, with the date of the last review.

What are the risks and limits?

Customer data. A template invites people to paste real information. ID numbers, bank accounts and health data do not go into personal or free accounts: use business accounts and replace names with codes when the task does not need them. Our guide to Colombia's Law 1581 covers this in detail; it is general information, not legal advice.
Made-up facts. A good prompt lowers the risk, it does not remove it. Prices, dates, references and regulations get checked against the source before anything is sent.
Nothing goes out unreviewed. The template produces a draft. A person is accountable for the message, especially for complaints, collections and anything involving money.
Tools change. As of October 2026, Claude, ChatGPT and Gemini all let you save instructions, but the names, limits and plans for those features change often. Check the current documentation.
A prompt will not fix a muddled process. If your team is unclear on the return policy, the AI will be too.

Is it worth the time?

A good template can be written and tested in an afternoon. After that, anyone can use it in seconds and the results look alike, which is what a company needs.

This blog works the same way at 77Rentals: the agents that research and write it follow written instructions on who they are writing for, what tone to use, which sources count and what they must not claim. It is the logic of these templates, applied to a whole process.