Work through customer communication, marketing claims, operations, and planning without exposing business or customer information.
Choose one real task, replace the brackets with safe context, test the result against a clear standard, and keep the person doing the work responsible for every claim and final choice.
What specific task and audience should the output serve?
What information is genuinely necessary—and what must stay out?
What source, standard, or expert will be used to check the answer?
What decision must remain with a qualified person?
What would make a non-AI approach safer, faster, or more appropriate?
The source guidance emphasizes protected information, lived and professional context, verification, accountability, and the limits of generated content.
U.S. Small Business Administration
SBA emphasizes human review, data protection, cybersecurity, intellectual property, transparency, and careful vendor selection.
It shapes the safe-input rules, review steps, business-value test, and customer-trust prompts.
Cybersecurity and Infrastructure Security Agency
CISA advises people not to put sensitive information into AI tools, warns about AI-enabled deception, and emphasizes account security and human expertise.
It supports the input restrictions, security controls, synthetic-media rules, training expectations, and incident reporting steps.
Sources and requirements can change. Check the publication date, read the original material, and confirm the rules that apply locally before making a consequential decision.
Use the task, audience, local conditions, constraints, output format, and review test that fit the work. The same prompt should not be copied into every classroom, business, campus, or community.
State the real purpose: serve a customer, improve an operation, communicate clearly, or test an idea with measurable value. Name the audience, desired outcome, and how the result will be used before writing instructions.
Do not enter customer records, employee files, credentials, payment data, contracts, tax information, trade secrets, or unreleased plans. Use fictional, public, aggregated, or properly approved information instead.
Ask the system to separate facts from suggestions, identify uncertainty, avoid invented sources, and state what information it would need to improve the answer.
Check important facts in original sources, apply professional and community knowledge, revise for accessibility and context, and keep a person accountable for use.
✓ Creating a first draft from public or fictional information
✓ Generating questions, alternatives, or structures for a person to evaluate
✓ Reformatting text after accuracy, context, and accessibility review
× Uploading private, confidential, protected, or proprietary information
× Requesting a final consequential decision or professional judgment
× Publishing an answer, citation, statistic, translation, or claim without verification
Replace brackets with safe, relevant context. Each template states what the person using it must still know, check, change, and own.
Create a respectful draft without exposing customer information.
1. Task: Draft a response to a fictional customer question about [product/service issue].
2. Context I can safely provide: Public policy, product facts, desired tone, and the resolution options an employee is authorized to offer.
3. Output: A short response, a list of facts that still need confirmation, and an escalation note.
4. Boundaries: Do not invent facts, quotations, statistics, people, policies, citations, or sources. Mark uncertainty and ask for missing information.
5. Human review: A knowledgeable employee verifies every promise, price, policy, and resolution before sending.
Stress-test a public claim before publication.
1. Task: Review this draft claim for clarity, evidence, likely misunderstanding, accessibility, and unsupported promises.
2. Context I can safely provide: The draft, intended audience, verified evidence, and channel—without customer or confidential information.
3. Output: A claim-by-claim table: supported, needs evidence, unclear, or remove; then a safer rewrite.
4. Boundaries: Do not invent facts, quotations, statistics, people, policies, citations, or sources. Mark uncertainty and ask for missing information.
5. Human review: The owner checks evidence and applicable advertising, industry, accessibility, and brand requirements.
Find bottlenecks without handing over the decision.
1. Task: Ask me one question at a time to map [workflow] and identify delays, rework, handoffs, and customer pain points.
2. Context I can safely provide: Roles, public or fictional process steps, current measures, and constraints.
3. Output: A current-state map, three hypotheses, measures for a 30-day test, risks, and a non-AI alternative.
4. Boundaries: Do not invent facts, quotations, statistics, people, policies, citations, or sources. Mark uncertainty and ask for missing information.
5. Human review: The process owner validates the map and chooses any pilot.
Convert non-sensitive notes into accountable follow-up.
1. Task: Turn these non-sensitive notes into a proposed action list.
2. Context I can safely provide: Attendees by role, decisions, deadlines, and open questions; exclude personnel and confidential business details.
3. Output: Owner, action, due date, dependency, decision still needed, and source note.
4. Boundaries: Do not invent facts, quotations, statistics, people, policies, citations, or sources. Mark uncertainty and ask for missing information.
5. Human review: Each named owner confirms the action and deadline.
Test one small-business work prompt against a real quality standard without using sensitive information.
Choose a low-stakes task and record what a good result must contain.
Adapt one prompt template using fictional or public context.
Run it once and mark accurate, useful, missing, unsupported, biased, or unclear parts.
Change one instruction and test again against the same standard.
Verify important claims outside the tool and decide whether AI added enough value to keep using.
Which improvement came from clearer instructions, and which still required human knowledge, source-checking, or a different tool?
Share the plan with the people responsible for the work and the people who may be affected. Confirm local rules, protect private information, and name the person who can answer questions or correct a problem.