A practical workshop for drafts you can check. Six original worksheets, a worked example, and a decision to keep, change or stop. No invented success stories.
Educational edition. Independent specialist review and reader testing remain pending. Hypothetical examples are not business evidence.
A bounded experiment for one recurring task—not a prediction about your job.
Name one task and the person who uses its output. Be more specific than “write reports”.
Describe the bottleneck: too many inputs, unclear instructions, repetitive drafting, or decisions that need expertise.
Choose one draft or organising step. Keep the final decision with a person.
Write three observable acceptance criteria and one stop condition.
Compare with your normal process, including setup, checking and correction time.
Your worksheet
No confidential information. Filling this sheet does not send data to an AI provider. Blank prompts can be copied separately.
Task and intended reader
Current process and bottleneck
Approved inputs (no confidential data)
One AI-assisted step
Three acceptance criteria
Reviewer, stop condition and fallback
Optional prompt · use only approved inputs
Help me test one recurring task. Ask what result is needed, what approved inputs exist, what mistakes matter, and who can check the output. Suggest one bounded draft step or say AI is unsuitable. Produce three acceptance criteria, a stop rule and a manual fallback. Do not invent facts, savings or a job-risk score. Do not request sensitive records.
Using this prompt in another service sends your inputs to that service. Check its settings first. It is not a tested guarantee of output quality.
Before you use the result
Can a reviewer check every important output against an approved input?
Is the test reversible and separate from live actions?
Would the experiment still be useful if the result were “do not use AI”?
A handover that explains what matters, what is unknown, and what must not happen.
State the outcome and audience. Specify what this task does not cover.
List the authoritative documents with version dates and owners.
Explain the operational reason for important rules, not just the rules themselves.
Separate facts, assumptions and unanswered questions.
Name the reviewer and actions requiring permission. Set a review date for the card.
Your worksheet
No confidential information. Filling this sheet does not send data to an AI provider. Blank prompts can be copied separately.
Outcome, audience and exclusions
Approved sources, versions and owner
Rules and why they exist
Known facts / assumptions / unknowns
Forbidden actions and approvals
Reviewer and next review date
Optional prompt · use only approved inputs
Turn my non-sensitive notes into a task handover. Organise outcome, approved sources, constraints, unknowns, prohibited actions and human approvals. Keep assumptions visibly separate from facts. Ask about missing rules instead of inventing them. A prompt is guidance, not a security control; identify which restrictions need actual permissions.
Using this prompt in another service sends your inputs to that service. Check its settings first. It is not a tested guarantee of output quality.
Before you use the result
Does an outsider know which document wins if sources conflict?
Are unknowns explicit rather than filled with plausible guesses?
Are permissions enforced outside the prompt where necessary?
A reusable acceptance-test sheet and practice spotting plausible mistakes.
Choose the actual decision the output supports.
For each important claim, name the source a reviewer can check.
Create a normal case, a missing-information case and an exception case.
Check for omissions, invented commitments, incorrect calculations and unsupported recommendations.
Reject or revise failures. Repeat the checks after revision and record what still needs a specialist.
Your worksheet
No confidential information. Filling this sheet does not send data to an AI provider. Blank prompts can be copied separately.
Output purpose and decision
Authoritative inputs
Normal case and expected result
Missing-information case and expected result
Exception case and expected result
Fail conditions and reviewer
Optional prompt · use only approved inputs
Review this draft against only the approved inputs I provide. Return a table: claim or omission, supporting input, pass/fail/unknown, and correction. Flag unsupported claims even if plausible. Do not manufacture citations. Do not call the work safe or compliant. Suggest tests for normal, missing-information and exception cases.
Using this prompt in another service sends your inputs to that service. Check its settings first. It is not a tested guarantee of output quality.
Before you use the result
Does every pass have a traceable reason?
Are omissions checked as well as sentences that are present?
Is the cost of a wrong answer reflected in who reviews it?
Small teams deciding whether to continue · 20-minute exercise
Pilot Decision Sheet
A keep, change or stop decision supported by your own observations.
Choose a fixed batch of representative, low-risk tasks. Record the usual time and quality.
Measure briefing, generation, checking and correction separately. Include setup and tool charges.
Record mistakes by consequence, not just count. Protect the trial from customers and live actions.
Ask staff whether the new process creates extra work elsewhere.
Compare observed quality and total effort. Keep, change or stop; choose a date to review the decision.
Your worksheet
No confidential information. Filling this sheet does not send data to an AI provider. Blank prompts can be copied separately.
Task batch and baseline
Setup time / per-task draft, check and correction time
Tool and maintenance costs
Failures and consequences
Staff/customer effects and reviewer
Keep / change / stop, reason and review date
Optional prompt · use only approved inputs
Help me compare a limited trial with its baseline. Use only my measured inputs. Include setup, review, correction and tool costs. Show the calculation, missing evidence and quality failures. Recommend keep, change or stop with conditions. Do not invent ROI, future savings or customer outcomes.
Using this prompt in another service sends your inputs to that service. Check its settings first. It is not a tested guarantee of output quality.
Before you use the result
Are baseline and assisted tasks comparable?
Is checking and rework included?
Would a serious quality failure override a time saving?
Small teams setting boundaries · 25-minute exercise
Small-Business AI-Use Starter Policy
An editable operational policy to discuss with your team—not a compliance certificate.
List approved tools, accounts, owners and permitted low-risk uses.
Create a never-enter list and clarify approved data settings.
Keep a named person responsible for factual checks and final publication.
Require separate approval for external messages, payments, deletions and access changes.
Tell staff how to report a mistake without blame. Set a review date and update after incidents.
Your worksheet
No confidential information. Filling this sheet does not send data to an AI provider. Blank prompts can be copied separately.
Policy owner, scope and review date
Approved tools and permitted uses
Never-enter information and data settings
Output review and disclosure rules
Permissions, payment checks and incidents
Training, applicable obligations and escalation contact
Optional prompt · use only approved inputs
Help draft an operational AI-use policy from my approved tools and low-risk uses. Include prohibited data, human review, actual permission controls, staff training, incident handling and a review date. Keep placeholders for missing facts. Flag legal questions for qualified review; do not certify compliance or assume the Privacy Act applies identically to every business.
Using this prompt in another service sends your inputs to that service. Check its settings first. It is not a tested guarantee of output quality.
Before you use the result
Can staff identify the approved account and never-enter list?
Are actual access controls separate from policy wording?
Have applicable obligations and exceptions been checked?
People depending on an AI service · 20-minute exercise
Provider Exit Checklist
A tested record of what you can recover, reuse and revoke.
Inventory documents, instructions, decision records, integrations and provider-held memory.
Attempt a non-sensitive export. Record its format and whether another tool can use it.
Separate your source records from model-specific summaries, embeddings and hidden state.
Rehearse revoking access and doing the task manually. Check retention and deletion terms.
Record verified, unverified or unavailable for each capability. Review after terms or pricing changes.
Your worksheet
No confidential information. Filling this sheet does not send data to an AI provider. Blank prompts can be copied separately.
Dependencies and owner
What exports / format / test date
What does not transfer
Access revocation and manual fallback
Retention, deletion and pricing questions
Verified / unverified / unavailable, next action
Optional prompt · use only approved inputs
Audit my dependency on this service using only evidence I provide. Separate documented export rights from capabilities actually tested. List what may not transfer, access revocation, retention questions, switching effort and a manual fallback. Use verified/unverified/unavailable. Do not invent contract terms, legal clauses or guarantees of portability.
Using this prompt in another service sends your inputs to that service. Check its settings first. It is not a tested guarantee of output quality.
Before you use the result
Have you attempted the export, not just read a marketing claim?
Can someone find the original rules without the provider?
Does the fallback work if the service is unavailable today?