In this guide
- What is a ChatGPT prompt?
- How to write a good prompt with the CADRE framework
- A universal prompt template to customize
- How to adapt the 25 prompts to your work
- Privacy: what data can you submit?
- Checklist before reusing an answer
- Sales and prospecting
- Marketing
- Customer support
- Operations
- Low-risk management and HR
- Frequently asked questions about ChatGPT prompts
- Official sources
The 30-second summary
A useful prompt states the context, requested action, permitted data, desired result, and evaluation criteria. The CADRE framework in this guide brings those five elements together. It makes the request clearer, but never guarantees an accurate response: a qualified person should verify every consequential output.
Key takeaways
- 01Start with one precise task instead of a long list of instructions.
- 02Provide only the information that is necessary and authorized by your organization.
- 03Define the output format and the criteria you will use to review it.
- 04Ask ChatGPT to flag missing information rather than fill gaps with assumptions.
- 05Treat every response as a draft that requires review, verification, and approval.
What is a ChatGPT prompt?
A prompt is the instruction or content you give ChatGPT to obtain a response. Depending on the available features, it may include text, an image, a file, or audio. It can request writing, analysis, content transformation, or a sequence of steps.
Specificity helps the model understand your intent, but it does not make the answer guaranteed truth. OpenAI recommends stating the task clearly, providing relevant context, defining the desired tone or format, and refining the request iteratively.
OpenAI Help Center: How do I create a good prompt for an AI model?
How to write a good prompt with the CADRE framework
CADRE is an editorial synthesis created by Devauras to make established prompting practices easier to apply. It is not an official OpenAI method or a scientifically guaranteed formula. Use it as a checklist, then adapt it to the task and your organization’s rules.
| Element | Question to ask | Example |
|---|---|---|
| C — Context | Who is speaking, to whom, and in what situation? | A B2B SME following up after a prospect demonstration |
| A — Action | What single task should be completed? | Write a follow-up email |
| D — Data | Which authorized information should be used? | Call notes, offer details, and stated objections |
| R — Result | What format, tone, and level of detail do you need? | 120 words, professional English, one call to action |
| E — Evaluation | How will you check the quality of the response? | No invented promises; facts limited to supplied notes |
OpenAI Help Center: How do I create a good prompt for an AI model?
A universal prompt template to customize
Copy this structure before selecting one of the specialized templates below. Replace every bracketed instruction and remove unnecessary lines. The more relevant your context is, the less decorative instruction you need.
- 01
Context — We are [type of organization]. The recipient is [audience]. The relevant situation is [context].
- 02
Action — Your main task is to [action verb + concrete outcome].
- 03
Data — Use only [authorized information or documents]. If information is missing, ask a question or flag it.
- 04
Result — Respond as [format], in [language and tone], using [length or structure].
- 05
Evaluation — Before answering, check [criteria]. Do not invent figures, sources, or commitments absent from the data.
How to adapt the 25 prompts to your work
- Replace the brackets — An uncustomized template tends to produce a generic response. Add the product, audience, objective, and constraints that genuinely affect the task.
- Provide a reference — When authorized, attach an approved example, tone guide, or source document and tell the model to stay within it.
- Work in stages — For a complex task, request an outline first, approve it, and then request the draft. This makes errors easier to correct before they spread.
- Separate creation from review — Request a first draft and then apply an explicit review checklist. A second model pass does not replace human validation.
- Keep decisions human — Do not use a prompt to automate a sensitive decision about a person, contract, health, legal matter, or finances.
OpenAI Help Center: How do I create a good prompt for an AI model? · EUR-Lex: Regulation (EU) 2024/1689 on artificial intelligence
Privacy: what data can you submit?
Before pasting a document into ChatGPT, check your internal policy, the workspace or plan being used, and the nature of the data. Minimize information, anonymize it where possible, and avoid trade secrets, credentials, personal data, or sensitive data in an unapproved consumer tool.
OpenAI states that data from its Business, Enterprise, and API offerings is not used to train its models by default. That provider policy does not, by itself, make your use compliant. Your organization remains responsible for its lawful basis, internal instructions, access controls, retention, and oversight of outputs.
France’s CNIL advises small businesses to establish usage rules, protect confidential information, and verify generated content. Belgium’s Data Protection Authority reiterates that data-protection principles apply to artificial-intelligence systems. The EU AI Act also introduces obligations that vary according to an organization’s role and the risk level of the system.
OpenAI: Enterprise privacy at OpenAI · CNIL: Using generative AI in small and medium-sized businesses · Belgian Data Protection Authority: Artificial intelligence and data protection · EUR-Lex: Regulation (EU) 2024/1689 on artificial intelligence
Checklist before reusing an answer
- Do factual claims match verifiable documents or sources?
- Do the figures, quotations, links, and proper names actually exist?
- Does the response follow every requested constraint and format?
- Does it contain personal, confidential, or unnecessary information?
- Does the text create an unauthorized commercial, legal, or financial promise?
- Has a qualified person reviewed the output before publication or decision?
- Does it respect brand voice, third-party rights, and internal policy?
- Does the use case require a specific GDPR or AI Act assessment?
CNIL: Using generative AI in small and medium-sized businesses · Belgian Data Protection Authority: Artificial intelligence and data protection · EUR-Lex: Regulation (EU) 2024/1689 on artificial intelligence
25 professional ChatGPT prompts to copy and adapt
Each template follows the spirit of CADRE. Replace the bracketed text, remove unauthorized data, and review the result before using it. These are editorial starting points and are not presented as having undergone empirical testing.
Sales and prospecting
Prepare commercial conversations without inventing information about prospects or promises about your offer.
01Prepare a discovery call
Context: I am preparing a first call with a [prospect type] in [industry]. We offer [offer]. Action: Build a discovery-call plan focused on the prospect’s needs. Data: Use only this public or authorized information: [information]. Result: Provide 10 open questions grouped by objectives, current process, problems, impact, and decision criteria. Add a natural transition between groups. Evaluation: Do not invent facts about the organization. Flag information that should be researched or requested.
02Turn notes into a CRM summary
Context: These are rough notes from a sales conversation: [anonymized notes]. Action: Turn them into a usable CRM summary. Data: Remain strictly faithful to the notes. Result: Use the headings Situation, Need, Impact, Objections, Mentioned decision-makers, Next step, and Missing information. Evaluation: Clearly distinguish facts, prospect statements, and hypotheses. Add no information that is absent.
03Write a follow-up email
Context: After [meeting type], I need to write to [recipient role]. Our brand voice is [tone]. Action: Draft a concise, useful follow-up email. Data: The confirmed points are [points]. The agreed next step is [step]. Result: Provide a subject line and an email of no more than 120 words with one call to action. Evaluation: Do not create a promise, discount, or deadline that is not in the data.
04Prepare responses to objections
Context: A prospect raised these objections: [objections]. Our offer genuinely covers [capabilities] and does not cover [limitations]. Action: Prepare honest responses for a salesperson. Data: Use only the capabilities and limitations provided. Result: For each objection, give an empathetic restatement, a clarifying question, a factual response, and a proposed next step. Evaluation: Avoid pressure tactics and flag any objection requiring input from a product, legal, or security expert.
05Personalize a re-engagement message
Context: I am following up with a [prospect type] after [event]. Here is the authorized history: [history]. Action: Write three personalized follow-up variants. Data: Use only the history and this value proposition: [proposition]. Result: One direct version, one advisory version, and one very short version; no more than 90 words each. Evaluation: Each version must give a concrete reason to reply, remain respectful, and avoid false urgency.
Marketing
Turn a vague idea into a reviewable brief or draft without allowing the model to invent evidence.
01Create a campaign brief
Context: [Company] is launching [offer] for [audience] through [channels]. The objective is [objective]. Action: Create a campaign brief ready for team review. Data: Use [positioning, permitted evidence, constraints, budget, or timeline]. Result: Structure the answer as objective, audience, customer tension, message, proof, channels, deliverables, timeline, and indicators. Evaluation: Flag missing data and invent no testimonial, customer result, or product benefit.
02Find editorial angles
Context: We want to cover [topic] for [audience] at the [awareness/consideration/decision] stage. Action: Propose 12 distinct, useful editorial angles. Data: Our verifiable expertise is [expertise]. Existing content is [list]. Result: For each angle, give a promise, likely search intent, three subsections, and one primary source to consult—without inventing its URL. Evaluation: Exclude repetitive, sensational, or unsupported angles.
03Rewrite for a persona
Context: Here is an approved source text: [text]. The new audience is [persona], who understands [level] and wants [goal]. Action: Rewrite the text for this audience without changing the facts. Data: Preserve the claims, limitations, and sources in the original. Result: [format and length], in a [tone] voice and vocabulary accessible to [persona]. Evaluation: Separately list any source sentence that appears ambiguous or unsupported instead of strengthening it.
04Build a content calendar
Context: We publish on [channels] for [audience]. This month’s objective is [objective], and our priority themes are [themes]. Action: Propose a four-week editorial calendar. Data: Maximum frequency [frequency], available resources [resources], important dates [dates]. Result: A table with date, channel, topic, angle, format, funnel stage, CTA, and owner to confirm. Evaluation: Balance educational and commercial topics, avoid repetition, and flag unresolved dependencies.
05Critique an advertising message
Context: This advertisement targets [audience] on [platform]: [ad copy]. Action: Analyze its clarity and propose an improvement. Data: The permitted promise is [promise], available evidence is [evidence], and brand constraints are [constraints]. Result: A five-criterion diagnosis followed by three variants with headline, body copy, and CTA. Evaluation: Add no figure, superlative, guarantee, or urgency that the data does not justify.
Customer support
Speed up response preparation while retaining human escalation for sensitive or uncertain cases.
01Classify customer requests
Context: I need to organize synthetic or anonymized customer requests. Our approved categories are [categories]. Action: Assign a category and indicative priority to each request. Data: Requests: [list]. Priority rules: [rules]. Result: A table with Request, Category, Proposed priority, Reason, and Escalation required. Evaluation: Infer no personal characteristics. Use “review required” when a rule does not support a conclusion. A human makes the final decision.
02Draft an empathetic response
Context: A customer wrote: [anonymized message]. The applicable policy is [approved excerpt]. Action: Prepare an empathetic, solution-oriented response. Data: Use only the message and policy provided. Result: No more than 150 words, acknowledging the issue, stating confirmed information, and giving the next step and timeline only if supplied. Evaluation: Admit no legal liability, promise no refund, and invent no deadline. Flag whether escalation is required.
03Create an FAQ from documentation
Context: I want to turn this approved documentation into an FAQ: [documentation]. Action: Create questions customers would genuinely ask and answer them. Data: The supplied documentation is the only permitted source. Result: 12 question-and-answer pairs, with a direct first sentence and no more than 80 words per answer. Evaluation: Add “to be confirmed by the team” when the source does not contain an answer.
04Summarize a conversation
Context: Here is an anonymized customer conversation: [conversation]. Action: Prepare a handoff summary for a human adviser. Data: Remain faithful to the conversation. Result: Use Request, Actions completed, Current status, Expressed sentiment, Commitments made, Next action, and Open questions. Evaluation: Separate facts from interpretations and briefly quote the relevant passage when a commitment was made.
05Define human escalation
Context: We are designing a support journey for [product/service]. Request types include [types]. Action: Propose an escalation matrix for review by support, legal, and security leads. Data: Available levels [levels], service hours [hours], existing rules [rules]. Result: A table with Signal, Proposed level, Team, Target time to confirm, Information to transfer, and Action forbidden to the bot. Evaluation: Always escalate threats, sensitive data, disputes, security risks, and requests outside policy.
Operations
Structure information and repetitive processes, then have their business owners validate the result.
01Produce meeting minutes
Context: Here are anonymized meeting notes: [notes]. Action: Create action-oriented meeting minutes. Data: Use only the supplied notes. Result: A five-line summary, confirmed decisions, actions with owner and due date when mentioned, open points, and risks. Evaluation: Mark “unassigned” or “not defined” when an owner or date is missing. Do not turn a proposal into a decision.
02Write an internal procedure
Context: The current process is described here: [approved description]. Its users are [roles]. Action: Turn it into a clear internal procedure. Data: Do not complete missing steps. Constraints: [tools, approvals, security]. Result: Objective, scope, prerequisites, numbered steps, controls, exceptions, owner, and version history. Evaluation: Flag every ambiguity and place human approval before any irreversible action.
03Compare vendors
Context: We are comparing [vendors] for [need]. Here is their verified information: [data]. Action: Organize a factual comparison without choosing for us. Data: Use only the supplied data and these weighted criteria: [criteria]. Result: Comparison matrix, missing information, risks, questions to ask, and a summary of trade-offs. Evaluation: Give no score when required data is missing and invent no feature or price.
04Identify automation candidates
Context: These are the current steps in a process: [steps], their frequencies [frequencies], and tools used [tools]. Action: Identify potentially automatable tasks for further investigation. Data: Security and approval constraints: [constraints]. Result: A table with Task, Trigger, Inputs, Output, Frequency, Risk, Human approval, and Complexity to confirm. Evaluation: Do not present automation as feasible without a technical audit; exclude sensitive or irreversible decisions without human control.
05Create a quality checklist
Context: The team must review [deliverable or process] before [publication/delivery]. The internal standard is [rules]. Action: Turn the rules into an operational checklist. Data: Use only the supplied rules. Result: Checks grouped by preparation, content, compliance, approval, and archiving, with an Expected evidence column. Evaluation: Each item must be observable and written as a yes/no question or value to enter. Flag contradictory rules.
Low-risk management and HR
Use ChatGPT to prepare supporting materials, never to make automatic decisions about hiring, performance, or a person’s future.
01Draft an inclusive job advertisement
Context: We are recruiting a [role]. The approved job description contains [responsibilities, essential skills, conditions]. Action: Draft a clear, inclusive job advertisement. Data: Use only the approved criteria. Result: Title, mission, responsibilities, essential skills, desirable skills, conditions, and recruitment stages. Evaluation: Remove unnecessarily gendered or exclusionary wording without changing essential requirements. Invent no salary or benefit.
02Build an onboarding plan
Context: A person is joining the team as [role]. Their approved 30-, 60-, and 90-day objectives are [objectives]. Action: Prepare an onboarding plan for manager approval. Data: Contacts [roles], tools [tools], mandatory training [training]. Result: A week-by-week table with objective, activity, resource, owner, and observable sign of progress. Evaluation: Infer nothing about the person and distinguish requirements, recommendations, and items still to confirm.
03Prepare training material
Context: I am training [audience] on [topic]. The approved learning objectives are [objectives], and the source documentation is [content]. Action: Propose materials for a [duration] session. Data: Remain faithful to the documentation. Result: Sequenced outline, key messages, examples, practical exercise, knowledge-check questions, and summary sheet. Evaluation: Identify points requiring a demonstration or validation by a subject-matter expert.
04Prepare a one-to-one agenda
Context: I am preparing a recurring one-to-one with a team member. The factual topics to cover are [topics]. Action: Create an agenda that supports a balanced conversation. Data: Use no personal information beyond the supplied, authorized details. Result: Opening, questions about work, blockers, support needed, development, decisions, and actions to confirm together. Evaluation: Avoid psychological diagnosis, personality judgment, or conclusions about performance.
05Rewrite an internal communication
Context: Here is an internal announcement whose substance is approved: [text]. It is for [audience], and the desired tone is [tone]. Action: Improve its clarity without changing the announced decision. Data: Preserve the exact dates, responsibilities, limitations, and contacts provided. Result: Subject, one-sentence summary, structured message, expected actions, and channel for questions. Evaluation: List any ambiguity that must be resolved before release and add no justification that was not supplied.
Frequently asked questions about ChatGPT prompts
What is a ChatGPT prompt?
It is the instruction or content given to ChatGPT to guide its response. A prompt can define the task, context, available data, expected format, and review criteria.
How do you write a good ChatGPT prompt?
Define one precise task, provide only relevant context, specify the expected format, and explain how the result should be checked. If the first answer is insufficient, refine the prompt iteratively.
Is a longer prompt always better?
No. Length does not guarantee quality. A prompt should contain the information that genuinely affects the task without contradictions or irrelevant context.
Should you give ChatGPT a role?
A role can clarify the perspective or tone, but it guarantees neither expertise nor accuracy. Pair it with a clear task, data, boundaries, and review criteria.
Can you submit confidential data to ChatGPT?
Do not submit confidential, personal, or strategic data without authorization and an approved framework. Check the plan, internal policy, access controls, and applicable duties, then minimize or anonymize data.
Does ChatGPT Business guarantee GDPR compliance?
No. The provider’s data-handling commitments are one part of an assessment, but the organization remains responsible for lawful use, security, transparency, and human oversight.
Can you trust a ChatGPT answer without review?
No. An answer may be incomplete or inaccurate. Verify facts, sources, figures, and consequences before publishing, sending, or using it in a decision.
Official sources
- 01How do I create a good prompt for an AI model? — OpenAI Help Center
- 02Enterprise privacy at OpenAI — OpenAI
- 03Using generative AI in small and medium-sized businesses — CNIL
- 04Artificial intelligence and data protection — Belgian Data Protection Authority
- 05Regulation (EU) 2024/1689 on artificial intelligence — EUR-Lex
Read next
Turn prompts into reliable AI workflows
Devauras helps you integrate AI with your tools, data, and human approvals so isolated experiments become controlled professional workflows.
Explore AI integration


