The 30-second answer
Lists of “magic prompts” are useless. What changes the quality of an answer is the context you provide and the success criterion you set.
A vague request backed by three relevant documents consistently beats a sophisticated formula with no material to work on. The skill to teach in training is therefore not writing incantations, but the art of gathering the right context and saying what a good result looks like.
Elements in a request that works: role, material, task, success criterion.
Project per recurring topic: that is what saves you from re-explaining the context in every conversation.
Of generative-AI use in French small businesses is for writing content (France Num 2025 barometer, the French government’s small-business digital program).
How to structure a request that gets results
The fourth element is the most profitable and the most neglected. Saying what a good result looks like costs one sentence and saves a round of rewriting. “One page maximum,” “factual tone, not salesy,” “do not invent any figures: if data is missing, leave a flagged blank” — instructions like these are worth more than any elaborate opening formula.
That last instruction deserves special mention in training: explicitly asking the assistant to flag gaps rather than fill them is the simplest protection against made-up claims.
Three real examples, before and after
The theory above makes more sense in practice. Here are three requests as they really show up in companies, and their corrected version. The difference never comes down to how sophisticated the formula is.
Case 1 — the client follow-up
BEFORE
“Write a follow-up email for a client who isn’t answering.”
AFTER
“Here is the complete email thread and the proposal sent on June 12 [attached documents]. The client hasn’t replied in three weeks. Write a short follow-up, from the CEO, that doesn’t ask for a reply again but brings something new. Six lines maximum. Don’t commit to any deadline: ours are not confirmed.”
What changes: the real material, a precise intention (“bring something, don’t ask again”), and an explicit prohibition that protects the company.
Case 2 — the meeting summary
BEFORE
“Summarize these meeting notes.”
AFTER
“From these meeting notes, output only two lists: the decisions made, and the open items with the name of the person who has to decide. Ignore the rest. If a decision is ambiguous in the text, list it as open rather than interpreting it.”
What changes: “summarize” is not a task, it’s a wish. Naming the output structure and the rule for handling ambiguity produces a document you can use as is.
Case 3 — the regulatory question
BEFORE
“Does this obligation apply to me?”
AFTER
“Here is the applicable text [attached document]. Our company has 40 employees, €6M in revenue and operates in sector X. Walk through the reasoning criterion by criterion and give your conclusion at the end. Only cite what is in the document provided; if a criterion isn’t covered there, say so instead of filling the gap.”
What changes: on a high-stakes topic, you provide the text rather than relying on the model’s memory, you ask for the reasoning so errors become visible, and you forbid filling gaps. This is the format to require on any regulatory topic.
Stop re-pasting the context every time
The biggest waste of time we see in companies is not poor wording: it’s pasting the same documents and the same explanations at the start of every conversation. Three features fix this, and they can be learned in one session.
Projects. One space per lasting topic — a client, a product, an initiative — where you upload the reference documents and standing instructions once and for all. Every conversation opened in that project inherits them. It’s the first habit to establish, and the Pro plan includes projects without the five-project cap of the Free plan.
Connectors. Instead of exporting and re-pasting, you connect the tools where your information already lives. This is what turns usage from an individual gadget into a shared work tool — and it’s also what needs the most attention on access rights.
Memory. It lets you pick up the thread of work already started. Useful, but worth knowing for another reason: what is remembered is kept over time, which should be spelled out in your AI usage policy.
The test that reveals a trained team
Look at how your employees’ conversations begin. If each one starts with a long copy-paste of context, the team is using a text engine. If they start straight with the request, because the context lives in a project, the team is using a work tool.
The verification protocol, in four steps
This is the part of training that actually protects the company, and the one most often skipped because it’s less spectacular than producing content. It takes twenty minutes to teach and it sticks.
Step 3 deserves a closer look, because it catches what checking against the source misses. A result can be technically accurate and absurd for your business. A 60% margin in a wholesale business, an eight-day collection period, an acquisition cost divided by ten: these values exist somewhere, but not at your company. No one in the business is better placed to spot it than the employee who knows the trade — as long as you’ve told them it’s their job.
Step 2 contains a trap that must be named explicitly in training: cross-checking a claim by asking the assistant again is not verification. It’s the most common mistake among intermediate users. A source is an official document, a vendor page, a legal text — not a second generated answer.
Five habits to build
Provide the material rather than describe it. Attaching the document beats summarizing it. Many disappointing answers come from context recounted from memory, and therefore impoverished.
Correct rather than start over. The beginner’s reflex is to rewrite the request from scratch when the answer isn’t right. Saying what’s wrong with the answer you got converges much faster — and keeps what was already good.
Ask for the reasoning on high-stakes topics. On a calculation, a trade-off or a contract review, requiring the intermediate steps makes errors visible. A bare conclusion can’t be verified.
Treat every figure and every reference as unverified. This is the rule that protects a company the most. An amount, a date, a legal article, the name of a standard: all of it gets checked at the source before it leaves the building. We learned this the hard way in our own editorial work.
Know when not to use it. Low-stakes, high-volume tasks are the right ground. A binding decision, a topic where no one on the team can judge the answer, or data your policy prohibits: those are cases for abstaining, and they must be named in training.
These habits apply to any generative assistant, not just Claude — see our comparison of AI assistants for business. To organize team-wide upskilling, see our team training program; for data governance, see what happens to your data.