Most sales teams that try AI stop halfway: they get copy that is correct, generic and lifeless, and they conclude it does not work. The problem is almost never the tool. It is that they brief it like a search engine instead of like a new rep who needs the business explained to them. This guide fixes that with 34 ready-to-use prompts, templates for Mexico and Paraguay, and the tricks you pick up in good training.
Short answer: A good sales prompt has six blocks: role, context with the reason behind it, labeled data, three to five real examples, an exact output format, and measurable success criteria. With that structure, the same 34 prompts cover the entire cycle: prospecting, first contact, discovery, proposal, objections, negotiation, follow-up, post-sale and win-back. Your company's style gets locked in through a shared project with your own examples and a list of banned words.
- What actually makes a sales prompt work?
- The anatomy of a good prompt: the six blocks
- Why do XML tags improve the output so much?
- Bad prompt vs. good prompt: the table worth printing
- Prospecting: five prompts for deciding who to call (1-5)
- First contact: six prompts for email and WhatsApp (6-11)
- Discovery: five prompts for asking better questions (12-16)
- Proposal and quoting: five prompts (17-21)
- Objections, negotiation and closing: seven prompts (22-28)
- Follow-up, post-sale and win-back: six prompts (29-34)
- Email and WhatsApp templates adapted to Mexico and Paraguay
- How do you build your company's “style” so Claude writes the way you do?
- How do you keep it from sounding like a robot? Common mistakes and how to fix them
- Tricks that are not in the manual
- What to do this week
What actually makes a sales prompt work?
A sales prompt works when you hand Claude the same thing you would hand a rep who joins your company today: who they are, who they are writing to, what information they have in front of them, what a good result looks like, and the standard you will review it against. When one of those five is missing, the copy comes out generic. And generic copy does not sell.
Anthropic's official documentation sums it up in an image worth taping to your monitor: think of Claude as a brilliant but brand-new employee who does not know your norms or your workflows. Capable and fast, but unaware that your sales cycle runs four months, or that the real buyer is the operations director and not the purchasing manager.
That is where the golden rule on its living page of prompting best practices comes from (accessed September 19, 2026): show your prompt to a colleague who knows little about the task. If they end up confused, so does Claude.
The second rule is less known and more powerful: explain the reason, not just the instruction. In the documentation itself, instead of "never use ellipses" you write "your answer will be read aloud by a text-to-speech engine that cannot pronounce them." Translated into sales terms: instead of "be brief," write "this gets read on a phone, standing up, between two meetings."
The third one almost nobody uses: success criteria. The official guide to developing tests asks for specific, measurable, achievable and relevant criteria, in categories such as task fidelity, tone and style, or privacy preservation. In a sales email that is as simple as: under 110 words, a single question, and zero empty adjectives.
The rest of this article applies those three rules to the nine stages of the sales cycle. If you do not have the overall framework yet, start with the phase-by-phase implementation guide.
The anatomy of a good prompt: the six blocks
A solid sales prompt has six blocks. When the output is bad, one of them is almost always missing.
- Role. One sentence defines who Claude is here. The documentation notes that a single line in the system message is enough to change behavior and tone.
- Context and reason. Who the customer is, what stage they are in, what happened before, and what this is for. The reason is the most commonly forgotten piece and the one that improves the output the most.
- Input data, labeled. Notes, history, price list: each one inside its own tag so they do not get mixed up.
- Examples. The official guide is explicit: include 3 to 5 examples for the best results, relevant, diverse and structured. Two emails that already won business are worth more than any description of tone.
- Output format. An email with a subject line and body, three numbered variants, a four-column table, a WhatsApp message under 60 words.
- Success criteria and limits. Banned words, a ban on inventing figures, and an instruction to flag the gaps.
Master template (copy it and keep it):
You are [ROLE: job title, industry, type of customer they sell to]. You work for [company, in one sentence].
<context> Customer: [industry, size, country]. Stage: [prospecting / discovery / proposal / negotiation / post-sale]. What happened before: [three lines]. Goal: [one sentence]. Why it matters: [the real reason]. </context>
<data>[notes, history, commercial terms]</data>
<examples><example>[a message of yours that actually worked]</example><example>[a different one]</example><example>[one from a hard case]</example></examples>
<format>[what you want to receive, in detail]</format>
<criteria>Maximum [X] words. A single question at the end. Banned: [list of words]. If you are missing a piece of information, write [MISSING: what data] instead of inventing it.</criteria>
That last criterion, the flagged gap, is the one that prevents the most trouble: without it, the model fills the hole with something plausible.
Why do XML tags improve the output so much?
Because they separate instruction from material. Without tags, Claude has to guess where your request ends and the transcript you pasted begins.
The documentation recommends tags such as instructions, context and input, with consistent names. You can use your own: call_notes, price_list, customer_history. What matters is that the name says what is inside and that you always use it the same way.
Order matters more than it seems
For long documents, from roughly 20,000 tokens on, there is a concrete rule: put the documents at the top, above the question, the instructions and the examples. The documentation states that putting the query at the end can improve quality by up to 30% in its tests. The same page recommends wrapping each document with its source and its content, and asking Claude to extract the relevant quotes first before answering.
Prompt for an RFP, a public tender or a long contract:
<document index="1"><source>Tender documents from [entity], [date]</source><content>[paste the full document]</content></document>
Before answering, extract into <quotes> the literal passages about mandatory technical requirements, deadlines, payment terms, warranties and grounds for disqualification. Number them.
Then, in <analysis>: what we meet as-is, what requires an exception, and what disqualifies us today, citing the number in each point. If something does not appear, write [NOT IN THE TENDER DOCUMENTS].
Anchoring the answer in quotes turns Claude into a fast, verifiable reader instead of a commentator.
Want a hand applying this?
Free 30-minute working session. Uniamos is a remote AI automation agency (Austin and Mexico City) working over video calls and WhatsApp.
Bad prompt vs. good prompt: the table worth printing
The difference between a bad prompt and a good one is almost never length. It is the number of decisions you leave unmade. Every decision you do not make, the model makes for you by averaging.
| What people usually write | What you should write instead | What changes |
|---|---|---|
| "Write me a sales email." | "You are the sales manager of a refrigeration distributor. Cold email of no more than 100 words to the maintenance manager of a 120-room hotel in Asuncion. Goal: a 15-minute call. A single question at the end." | Defines role, recipient, length, goal and close. It stops being a piece of text and becomes a tool. |
| "Make it more professional." | "Rewrite it with sentences under 20 words, no evaluative adjectives, in the second person, and without the word solution." | "Professional" means something different to everyone. Three concrete rules are verifiable. |
| "Summarize this call." | "Summarize it in five bullets: stated pain, budget mentioned, who decides, main objection, and commitment with a date. If something was not mentioned, write [NOT DISCUSSED]." | Fixes the structure and forbids filling gaps. |
| "Give me ideas to push back on we're too expensive." | "Three different responses: one that reframes to total cost, one that reduces scope, and one that asks for a piece of information before answering. None of them can promise a discount." | Three real angles instead of three ways of saying the same thing, with a clear limit. |
| "Make it shorter." | "Cut it to 70 words, removing adjectives first, then courtesies, and only at the end content. Tell me what you removed." | Defines the order of the cuts, which is what separates a good short text from a mutilated one. |
| "What do you think of this proposal?" | "Read it as the customer's CFO, who was in none of the meetings. Tell me the five uncomfortable questions you would ask and where you would stop reading." | Changing the reader's role is the fastest way to find the holes in a document. |
If you take away only one thing: replace the adjectives in your prompts with numbers or with rules. "Short" is not an instruction; "70 words" is.
Prospecting: five prompts for deciding who to call (1-5)
In prospecting the goal is not to write, it is to decide. These five prompts turn a long list into a short list with a rationale.
1. Ideal customer profile. I am giving you our 20 most profitable customers over 24 months with industry, size, country, channel and margin, and after that the ones we lost. Return three profiles with a short name, the signals that are observable from outside, and one signal of a bad fit. Do not invent patterns that are not supported by at least three cases.
2. Buying signals in the industry. We sell [what] to [type of company] in Mexico and Paraguay. Give me 12 public signals that a company like that might need us within six months: hiring, construction, openings, regulatory changes, public tenders. For each one: where you can observe it, how often to check it, and an opening line that uses it without sounding like surveillance.
3. Account research before calling. Find recent public information about [company, city, country]: what they do, size, news from the past 12 months, the people in charge of the area we care about, and three hypotheses about their problem. Half a page, each fact with a source and a date. Whatever you cannot find, mark as [NO DATA]. I am going to use this on a real call.
4. Prioritizing a long list. I am giving you 200 trade show contacts with company, job title, country and the rep's note. Sort them into A (call within 48 hours), B (email within 7 days) and C (newsletter), applying in this order: fit with the ideal profile, seniority, concrete pain in the note, and timeline. A table with contact, category, deciding criterion and a suggested opening line.
5. Qualifier for the booth. Design a five-question qualifier to use standing up at a trade show, in under 90 seconds. Each with three closed answers and their score, and a final threshold that separates A, B and C. Write them the way a person would say them, not like a form.
Why qualify at the moment of capture: the CEIR report How Exhibitors Can Improve Lead Quality and Sales Conversion (2019 edition, with data from 2015 and 2018, public PDF) reports that only 39% of trade show leads turn out to be qualified by the exhibitor's own criteria, and that fewer than 10% of exhibitors use custom qualifiers. The problem is not capturing more: it is qualifying while you capture, as we cover in the first-72-hours protocol.
First contact: six prompts for email and WhatsApp (6-11)
First contact is won on relevance and brevity. If you do not ask for them explicitly, they do not show up.
6. A 90-word cold email. Cold email of no more than 90 words to the [job title] at [company, industry, country]. Context: [two lines]. Goal: a 15-minute video call. First line about the recipient, not about us. No "I hope this email finds you well." A single concrete benefit. A single yes-or-no question at the end. No attachments and no use of the word "solution."
7. Six subject lines to test. Give me six subject lines in three pairs: two direct, two questions, and two with a verifiable reference to the company. Maximum 45 characters, no shouting caps and no emojis, and none of them promises something the body does not deliver. Next to each, one line on what kind of recipient would open it and what its risk is.
8. First WhatsApp message after consent. First message to a contact who gave us their number at [trade show / form / call] and agreed to be contacted. Maximum 45 words: person and company in the first line, where we met in the second, closed question at the end. No emojis and no exclamation marks. Also give me a version two notches more informal.
9. Voice note script. Turn this message into a script for a 25-to-30-second voice note: short sentences, spoken language, no complicated numbers, with a slash where it makes sense to pause.
10. Following up on a trade show contact within 72 hours. Contact captured at [trade show] on [date]. The rep's note, verbatim: "[paste it even if it is badly written]." Write an email of no more than 100 words and a WhatsApp message of no more than 45, both with a concrete reference to the note and ending with two specific time slots, not with "whenever works for you."
11. The same idea, two countries. Adapt this message for a recipient in [city in Mexico] and another in [city in Paraguay], with the same content and length. Adjust the form of address, the time reference, the currency, and the industry terms that differ. At the end, tell me in two lines what you changed and why, so I can validate it with the local team.
Asking it to explain what it changed lets you fix the prompt instead of fixing every message. If WhatsApp is your main channel, first look at how to set up the circuit with the official platform: business-initiated messages require an approved template and prior consent.
Discovery: five prompts for asking better questions (12-16)
Discovery is where AI adds the most and gets used the least. Not because it writes better questions than you do, but because it remembers all the ones you forget under pressure.
12. Industry-specific script. A script for a first 20-minute call with the [job title] at a [type of company] in [country]. Eight open questions, from least to most probing. For each one: the follow-up question if the answer is vague, and the signal that we are talking to the wrong person. None of them can be answered with yes or no.
13. Preparing the call with the history in front of you. <history>[emails, notes, tickets]</history> I have a call in 30 minutes. Give me the three facts I cannot forget, what we promised and did not deliver, the question that went unanswered, and two topics worth avoiding today. All of it has to come from the history; if something is not there, say so.
14. From loose notes to a usable summary. <notes>[paste them as they are, with typos and abbreviations]</notes> Turn this into five bullets (stated pain, budget mentioned, who decides, main objection, commitment with a date), three tasks with an owner and a deadline, and an 80-word follow-up email. If something was not mentioned, write [NOT DISCUSSED].
15. What was not said. Read these notes and tell me what is missing: what part of the business we did not touch, which member of the committee did not show up, what risk we did not ask about, and what information we would need to quote without assumptions. Order it by how likely it is to kill the deal, with the exact question to ask next time.
16. From the conversation to the CRM. Using these notes, fill in a table: stage, estimated amount, likely close date, competitor mentioned, reason to buy in the customer's own words, next step and its date. Use only these allowed values: [paste your stages]. If a field cannot be completed, leave it blank and note it as information to obtain.
Prompt 16 solves the oldest problem any sales team has: an incomplete CRM. When filling it in takes two minutes instead of twenty, it gets filled in. And the natural place to share that knowledge is a project, as we explain in the Projects guide.
Proposal and quoting: five prompts (17-21)
A proposal is not judged by what it says about you, but by how easy it is for the customer to defend it in front of someone who was in none of the meetings.
17. Structure based on discovery. <discovery>[summary of the calls]</discovery> Propose which sections, in what order, and how many pages, for a reader who is a [job title] and has ten minutes. Justify each section with a line from the discovery. If a standard section is not justified by anything the customer said, remove it and tell me why.
18. Scope and assumptions. Write this section as three lists: what is included, what is not included, and what we assume the customer will provide (data, access, people, deadlines). Every line verifiable: no "ongoing support" without stating hours and channel. Where something still needs defining, write [TO DEFINE] and move on.
19. From feature to consequence. I am giving you eight features exactly as the technical team writes them. Turn each one into a consequence statement for a [job title], using the customer's own vocabulary. Table: feature, what it means for them, how they would verify it. If one has no clear consequence for that role, flag it as irrelevant.
20. Read it as the person who says no. Read this proposal as the customer's CFO, who took part in none of the meetings and receives it as a forward. Tell me the five uncomfortable questions you would ask, the paragraph where you would stop reading, the three numbers you would ask us to justify, and what argument you would use to postpone it three months. Be harsh.
21. Executive summary and phone version. From this proposal, give me a one-page executive summary with problem, solution, investment, timeline and next step, plus a 50-word WhatsApp message the contact can forward to their boss. The same figures as the document; if one is still pending, leave it out of both.
Objections, negotiation and closing: seven prompts (22-28)
Objections do not get improvised well. You prepare them once, write them down, test them and correct them.
22. Your company's objection bank. I am giving you 15 emails and notes from deals lost over 12 months. Extract the real objections in the customer's exact words and group them by family. For each one: how many times it appeared, at what stage, what we answered, and with what result. Flag the two that appear latest in the cycle, which are the most expensive ones.
23. "You're too expensive." Context: [proposal, alternatives the customer is looking at, margin I can move]. Three different responses, not three versions of the same one: one that reframes to total cost and risk, one that reduces scope while keeping the unit price, and one that asks for a piece of information before answering. Maximum 60 words each. None of them promises a discount.
24. "We'll look at it next year." Three responses: one that quantifies the cost of waiting using the customer's own numbers, one that proposes a first step that fits this fiscal year's budget, and one that accepts the delay while setting a commitment with a date. The third one without sounding like surrender or like pressure.
25. "We already work with another supplier." Five questions to understand the real level of satisfaction with the current supplier without criticizing them once, and a closing line that leaves the door open to being a second option. No comparing, no naming the competition, and no implying they are being poorly served.
26. Preparing the negotiation. These are our terms and this is the margin: [paste]. Build my negotiation sheet: what I can concede and in what order, what I ask for in exchange for each concession, what my walk-away point is, and three signals that the customer has already decided and is only hunting for a discount. Order the concessions from lowest to highest cost for us.
27. Role-play with a tough buyer. You are going to play the [job title] at [company]: skeptical, with a current supplier, and short on time. Short sentences and some impatience. Do not make it easy for me. I will start. After ten exchanges, drop the role and tell me my three worst answers and the question that would have left me with nowhere to go.
28. Closing email with options. Three clear options (reduced scope, full scope, full scope with staged payments), each in two lines with price and timeline, and a single final question: which one do we move forward with. Maximum 120 words. No summarizing the proposal again and no artificial urgency.
Follow-up, post-sale and win-back: six prompts (29-34)
This is the money almost nobody picks up. A team that prospects well and follows up badly works twice as hard to sell the same amount.
29. Staged follow-up plan. For a sales cycle of [X] months, design a 90-day plan with concrete dates. For each touch: channel, reason (what I am giving them that day, not "checking in"), maximum length, and what I do if they do not reply. If on a given day there is nothing to offer, propose not writing and tell me why.
30. The follow-up that actually gives something. A contact who has not replied for three weeks. Here is what I know about them and their company: [paste]. Three messages of no more than 50 words: one that shares something useful and concrete, one that offers to close the topic gracefully if it no longer applies, and one that switches to a different person inside the same company. None of them starts with "just checking in to see if."
31. New customer onboarding. We closed with [customer]. Build the first 30 days: what we send on day 1, what we review on day 7, what we measure on day 30, who does each thing, and what we need from them with a name and a date. Also write the day 1 email, 100 words, covering what was agreed, without reading like a contract.
32. Early churn signals. I am giving you six months of history: tickets, purchases, emails, meetings. Tell me whether you see churn risk signals, classified as high, medium and low, with the data point behind each one and the action I would take this week. If there are no clear signals, say so: I prefer an honest "I do not see anything" to an invented list.
33. Reactivating dormant customers. A list of 60 customers who have not bought in more than 12 months, with what they used to buy, how much, and when they stopped. Group them into three groups by the most likely reason for the drop-off and write a different message for each group, maximum 60 words. None of them can say "it has been a while since we heard from you."
34. Weekly pipeline report. <pipeline>[CRM export]</pipeline> A half-page report: what moved, what stalled, what came in, what was lost and why. Then, three decisions management has to make this week. No adjectives. If a deal has been in the same stage for more than [X] days, flag it even if the rep says it is going well.
Want a hand applying this?
Free 30-minute working session. Uniamos is a remote AI automation agency (Austin and Mexico City) working over video calls and WhatsApp.
Email and WhatsApp templates adapted to Mexico and Paraguay
The prompts above generate copy; these are ready-written templates to paste and adjust. The difference between the two markets is not the language, it is the form of address and how much weight WhatsApp carries compared with email.
| Practical detail | Mexico | Paraguay |
|---|---|---|
| Form of address at first B2B contact | Formal usted in writing; the informal tuteo appears when the other side proposes it | Formal usted in writing; in the closer relationship that follows, voseo is common, so decide it with your local team |
| Channel that gets the most replies after a trade show | WhatsApp to confirm, email for the document | WhatsApp for almost the whole thread; email works as the formal record |
| What the commercial document is called | Cotización (quote) | Presupuesto or cotización, depending on the industry |
| Receipt the finance department will ask for | Electronic invoice (CFDI) | Electronic invoice from the national system |
| Currency in the proposal | Pesos, and dollars if the input is imported; always state which one | Guaraníes, often with a dollar reference; state which one and at what exchange rate |
| Sign-off that sounds natural | "Quedo atento" / "Me avisas" | "Quedo a la espera" / "Aguardo tu confirmación" |
These are practical conventions of etiquette, not hard rules: confirm them with your local team before locking them into a template.
The four templates below stay in Spanish on purpose. They are the text you send to a Spanish-speaking customer, so translating them would defeat the point; the headings and the notes around them are in English so you know what each one is for.
Cold email (Mexico version) - sample text in Spanish, ready to send
Asunto: Paro de línea en [mes] — una pregunta
Buenos días, [cargo]: vi que [empresa] abrió la planta de [ciudad] este año. En operaciones de ese tamaño el mantenimiento correctivo suele llevarse más horas de las previstas durante los primeros meses.
Trabajamos con [tipo de empresa] en [región] justamente en esa etapa. No le voy a mandar un catálogo.
¿Tiene 15 minutos el jueves o el viernes por la mañana?
Saludos, [nombre] — [empresa] — [teléfono]
Cold email (Paraguay version) - sample text in Spanish, ready to send
Asunto: Consulta breve sobre [tema concreto]
Buenos días, [cargo]: le escribo porque vimos que [empresa] está [hecho verificable: ampliando planta, abriendo sucursal, exportando a X].
Acompañamos a empresas de [sector] en Paraguay en ese momento concreto, sobre todo en [problema específico]. Se lo muestro en 15 minutos, sin presentación institucional.
¿Le sirve el martes a las 9 o el miércoles a las 8?
Quedo a la espera. [nombre] — [empresa] — [teléfono]
WhatsApp after a trade show and quote follow-up - sample text in Spanish, ready to send
Primer contacto: Buen día, [nombre]. Soy [nombre] de [empresa], nos conocimos ayer en el stand de [feria] y quedamos en mandarle el dato de [tema que se habló]. Se lo paso acá: [una línea útil de verdad]. ¿Le sirve una llamada corta el jueves a las 10 u 11?
Seguimiento de cotización: [Nombre], buen día. Le escribo por la cotización del [fecha]. Dos puntos suelen definir la decisión: el plazo de entrega y quién absorbe [concepto]. 1) [opción A, precio, plazo] · 2) [opción B, precio, plazo]. ¿Avanzamos con alguna o le preparo una tercera?
Dormant customer win-back - sample text in Spanish, ready to send
[Nombre], buen día. Soy [nombre] de [empresa]. Su última compra fue [producto], en [mes y año].
Cambiamos [algo concreto que le habría resuelto el motivo por el que dejo de comprar]. Si el motivo fue otro, me sirve saberlo igual.
¿Le mando el detalle por acá o prefiere que lo llame?
One warning worth not skipping: sending business-initiated messages through the WhatsApp Business platform requires prior consent and approved templates, and that consent is collected on the booth form or the website form, at the moment of capture. This is not legal advice: how that data is handled is governed by each country's regulations and is worth checking. For Paraguay, see customer data and Law 7593.
How do you build your company's “style” so Claude writes the way you do?
By building three things, in this order: a bank of real examples, a permanent instruction block, and a list of bans. Nothing technical is required: the sales manager can set it up in an afternoon.
1. The example bank: 10 to 15 pieces of copy that already worked
Go through your sent mail for the messages that got a meeting, unblocked a deal, or rescued an angry customer, anonymize them and keep them together. Having 15 in the bank lets you pick the 3 to 5 that match each situation. Include at least two uncomfortable ones: a piece of bad news and a reply to an upset customer. Those are the ones that define a tone best.
2. The permanent instruction block
It lives in the instructions of a shared project. Claude projects have their own knowledge base and instructions, and on paid plans they enable retrieval-augmented generation, which expands capacity by up to ten times as you approach the context limit. On Team and Enterprise the project is shared with specific members or with the whole organization: that turns a personal trick into a company standard.
Project instructions for "Company sales style":
You write on behalf of [company], which sells [what] to [whom] in Mexico and Paraguay.
How we write: sentences under 20 words, active voice, second person, numbers before adjectives, one topic per message and one question, always at the end.
Form of address: formal usted at first written contact; informal tuteo only if the customer starts it; in internal messages, always informal.
We never use: solución integral, sinergia, ecosistema, disruptivo, potenciar, valor agregado, en la era de, hoy más que nunca (in English copy, the equivalents: end-to-end solution, synergy, ecosystem, disruptive, leverage, added value, in the era of, now more than ever).
We never do: open with "I hope this email finds you well"; use exclamation marks at first contact; promise timelines before finance confirms them; share other customers' results.
If a piece of information is missing: write [MISSING: what data] and move on. Do not estimate, do not invent, do not round.
Before delivering: check against these rules and tell me in one line which one you came closest to breaking.
3. Turning it into a reusable capability
When that block stops changing every week, package it. Skills are folders of instructions that Claude loads when they are needed; they are written in Markdown, they are available on every plan, and they require code execution to be turned on. On Team and Enterprise, owners can provision them to the whole organization, and the official guide can be followed without knowing how to program.
A note in the interest of honesty: if your account offers any writing-styles option, use it; here we solve it with projects and skills because those are the mechanisms publicly documented as of September 2026. And a practical rule: the style gets reviewed once a quarter. If every rep tweaks it, it stops being a standard.
How do you keep it from sounding like a robot? Common mistakes and how to fix them
It sounds like a robot when the copy is correct but risks nothing: it mentions nothing specific, it has no rhythm, and it does not say who wrote it. These ten mistakes cover almost every case.
- Zero specificity. If it would work just as well for another forty companies, it is not personalization. Fix: the first line has to contain a verifiable fact about the recipient, and if there is none, the prompt tells you instead of inventing one.
- Empty courtesy opener. "I hope this email finds you well" burns the most valuable line. Fix: an explicit ban in the project instructions.
- Three ideas in one message. The customer replies to none of them. Fix: one topic, one question.
- Adjectives instead of data. "Significant improvement" means nothing. Fix: every benefit carries a number or gets deleted.
- Every sentence the same length. It is the biggest giveaway of generated text. Fix: ask for variation explicitly, for example alternating one long sentence with two short ones.
- Vocabulary your team does not use. Fix: the banned-words list, which is the rule that improves the output fastest.
- Ambiguous close. "Looking forward to your thoughts" asks for nothing. Fix: two concrete time slots or a yes-or-no question.
- Invented figures. The most serious mistake, because it destroys trust. Fix: the flagged-gap rule, with no exceptions.
- Too much enthusiasm. Exclamation marks, caps and emojis at first B2B contact work against you. Fix: ban them at the start and allow them if the customer uses them.
- Sending without reading. The one mistake no instruction fixes. Fix: nothing goes out without a person reading it in full.
Two thirty-second tests before you send
The read-aloud test. If you stumble, so does the customer. Anything you would not say on the phone gets deleted.
The colleague test. Show the message to someone on the team who does not know the account. If they have to ask you something to understand it, context is missing. If they say "you could send this to anyone," specificity is missing.
And a management principle: responsibility for the message always belongs to whoever signs it. That is worth writing into your internal policy, along with which customer data can be pasted into a conversation, as we cover in the AI use policy and governance guide.
Tricks that are not in the manual
The shortcuts that show up in good training and not in the documentation. All of them apply today and none of them needs special permissions.
- Start the prompt from the end. Write first what you want the result to look like and fill upward. It forces you to decide the format before you ramble.
- Ask for three different versions, not three variants. "Three options" returns the same text with synonyms. "One direct, one consultative and one provocative" returns three real approaches.
- The ban trick. Ten banned words improve the tone more than three paragraphs describing how you want it to sound. Start with: solución integral, sinergia, disruptivo, potenciar, valor agregado (in English: end-to-end solution, synergy, disruptive, leverage, added value).
- Split drafting and critique into two steps. The official documentation calls it self-correction chaining: draft, review against criteria, refine. Asking for all of it at once produces mediocre copy that praises itself.
- Change the reader's role, not the writer's. "Read it as the CFO who was not in the meeting" finds more holes than any "be more critical."
- Paste your ugly notes as they are. Cleaning them up takes ten minutes and does not improve the result; what you should do is remove personal data that does not need to be there.
- Long document, always at the top. Documents first, question at the end: Anthropic reports improvements of up to 30% in answer quality in its tests.
- Demand quotes before conclusions. Literal quotes first, analysis afterward. It turns an opinion into something you can verify in thirty seconds.
- Save the prompts that work, not the outputs. The output is good once; the prompt is good all year.
- Always name your tags the same way. If today you use call_notes and tomorrow notes, your prompts stop being reusable between colleagues.
- One live conversation per large account. On important accounts, stay in the same thread: the accumulated context saves you from pasting the history every time. On small accounts, one chat per task.
- Ask for the metaprompt when you get stuck. Ask it to rewrite the prompt, explaining what you wanted and what you got. Anthropic also publishes a prompt generator, an interactive tutorial and the Claude Cookbooks, all free.
And one management trick worth all the others: measure a single thing at the start, for example the minutes between the end of a call and the follow-up going out. If that number does not drop, the problem is not the prompt.
What to do this week
Five days, one hour a day, with a single person leading. At the end of the week you have a standard, not an experiment.
- Monday. Gather 15 sent messages that actually worked: three first contacts, three follow-ups, three proposals, three objections and three post-sale. Anonymize them.
- Tuesday. Copy the project instructions template, adapt it, and add your list of ten banned words. Create the project and upload the 15 examples there.
- Wednesday. Pick five of the 34 prompts that match what your team does most often each day and run them on real cases from last week, whose outcome you already know. Compare.
- Thursday. Two of the five will come out badly. Fix them by adding constraints (words, format, bans), not by rewriting the output by hand.
- Friday. A 45-minute session with the team: the five prompts, the banned-words list, and the two thirty-second tests. Decide who maintains the project and schedule the quarterly review.
Control checklist
- Does every prompt have a role, context, reason, format and success criteria?
- Do they all include the rule about flagging gaps instead of inventing data?
- Is there a banned-words list, and does the whole team know it?
- Are the 15 examples anonymized and in a shared project?
- Is there a written rule about which customer data does not get pasted into a conversation?
- Does someone read every message before it goes out?
- Is there a single agreed metric for knowing whether this is working?
- Is the style review scheduled three months out?
To turn it into a formal program, the next step is the eight-week program, and to know whether it pays off, the adoption and return guide. The courses and tutorials at Claude Academy are free and serve as common ground; bear in mind that they do not grant a credential and there is no official Anthropic certification program. If you would rather do it with support, at Uniamos we deliver this training using your own team's cases.
Frequently asked questions
Which Claude plan does a sales team need?
For individual work, a paid personal plan is enough. For a team, the relevant difference is not features but data policy: Anthropic documents that in its commercial products (Team, Enterprise and API) your inputs and outputs are not used to train models by default, whereas on consumer plans it depends on a user setting. As of September 2026 the standard Team seat costs 20 dollars per month billed annually or 25 per month, with a two-seat minimum; confirm on their site.
How many examples should I include in a prompt?
The official documentation recommends between 3 and 5 examples, relevant, diverse (covering edge cases) and structured, each one inside its own tag. In sales, the practical approach is to keep a bank of 10 to 15 real examples and pick the ones that match each situation.
Are XML tags necessary or is that overkill?
They become necessary as soon as you paste long material: notes, histories, tender documents or price lists. Their job is to separate the instruction from the material so the two do not get confused. Practical rule: if your prompt has more than one distinct block of information, tag them.
Is it safe to paste customer data into a conversation?
It depends on the plan and on the internal policy you have written down. In the commercial products, Anthropic documents that by default it does not train on the content, and Team and Enterprise owners can turn off feedback submission in the organization settings, which is the route by which a conversation could be retained. Even so, minimize: paste what is necessary and nothing more. This is not legal advice.
How do I stop it from inventing figures in a proposal?
With an explicit instruction in every prompt: if a piece of information is missing, write a marker such as [MISSING: what data] instead of filling it in. In long documents, ask for the literal quotes first and the analysis afterward. And a team rule: no figure goes out without a person verifying it.
Does this work for WhatsApp or only for email?
For both, but WhatsApp demands more discipline: 45 to 60 words, no emojis at first contact, and a single closed question at the end. Remember that business-initiated messages through the WhatsApp Business platform require prior consent and approved templates.
Which model is best for sales tasks?
According to Anthropic's official models page, accessed on September 19, 2026, the general recommendation is to start with Claude Opus 5 for most workloads, with Sonnet 5 as the best combination of speed and intelligence and Haiku 4.5 as the fastest. For writing emails the difference is small; it shows up in long documents.
How long does it take a team to notice the difference?
The writing and summarizing side shows up in the first week, because the change is immediate and verifiable. The process side, such as a current CRM or staged follow-up, takes four to eight weeks because it means changing habits. That is an estimate, not a measured figure: set your own baseline metric before you start.
Is there an official Anthropic certification for salespeople?
No. Claude Academy offers free courses and tutorials, but it does not grant a credential and there is no professional certification program announced as of September 2026. If a provider offers you an official Anthropic certification, they are not describing something that exists.
Related articles
- How to implement Claude in your company, phase by phase
- Projects: your company's shared brain
- Claude use cases by department in an SMB
- A B2B lead generation system powered by AI
- Post-trade-show follow-up in 72 hours with WhatsApp
- An eight-week AI training program for teams
- AI agent on WhatsApp Business: the definitive guide
- More from the blog