Editor-selected existing cases and prompts, separate from AI news. Selection dates use Hong Kong time.
Oct 24: five existing cases rechecked, not new or tested. Web drafts, not source agents/add-ins. Gemini web lists HK/MO/TW and Traditional Chinese; ChatGPT/Claude list Taiwan, not HK/MO. Free/basic text access; check image access and tool limits. Local paid prices and Chinese quality unknown. Use shareable data and human checks; no orders, deployment, overwriting or hardware control.
Supply menu text, party size and budget to draft an order using listed items, with unclear prices and special requests flagged for staff.
Prompt
Using [menu text/preferences/party size/budget and currency], draft an order list limited to listed items. Preserve original names and price evidence. Flag unclear text, unlisted charges and modifications for staff confirmation. Do not determine allergen safety, invent discounts, place orders or pay.
Evidence limits and cautions
A model cannot confirm live availability, prices or allergens; staff verification is needed.
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
- Do not treat model output as an allergen or health guarantee.
- Do not invent dishes, discounts or ingredients.
Original source: Restaurant staff observations of ChatGPT-assisted orders
Supply a technical draft, software versions and real test records to map claims to evidence and suggest small, reviewable corrections.
Prompt
Review [technical draft] against [versioned official docs] and [actual test records]. Build a claim-check table with evidence, questions and minimal suggested changes. Say when code was not run. Preserve [audience/tone/length requirements] and genuine personal experience; do not publish or fabricate results.
Evidence limits and cautions
No error-detection rate is provided and individual suggestions may be wrong.
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
- Reading code is not evidence it ran successfully.
- Do not invent the author’s tests or experiences.
Original source: Automating DevRel: How I Use Gemini CLI and Gemini 3 to Catch Bugs in My Blog Posts
Give a text AI a test site and approved link destinations to draft a link, keyboard and layout checklist for human review. This is a drafting exercise; the source used GitHub Copilot.
Prompt
Using [test site/link inventory/approved destinations/device range], draft link-change acceptance checks. Match visible labels to actual destinations, internal or external navigation, keyboard focus and narrow-screen wrapping. Ask about unknown destinations rather than inventing URLs. List existing navigation behavior that must remain unchanged and regression checks. Deliver a plan only; do not merge or deploy.
Evidence limits and cautions
The article supplies no comprehensive accessibility or cross-device acceptance results.
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
- Do not impose one website’s spacing values on unrelated designs.
- Do not merge or publish directly.
Original source: My first impressions of GitHub Copilot coding agent
Supply a fictional manual and editing rules to draft a change list with original locations, suggested wording and reasons for individual review.
Prompt
Using [fictional internal manual/editing goals/protected passages/formatting rules], propose individually reviewable revisions. Include original location, suggested wording and rationale for each, plus numbering and cross-reference checks. Provide a review draft only; do not assess legal effect, accept changes or overwrite the original.
Evidence limits and cautions
One trial does not establish reliable handling of complex formatting, cross-references or every edit.
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
- Do not assess contractual rights or provide legal advice.
- Do not accept all changes or overwrite the original.
Original source: Claude for Word Gives Lawyers Native Tracked Changes
Supply safe household-item photos and local rules to draft material classifications, preparation requirements and questions to verify.
Prompt
Draft categories for [safe household-object photos] using [region/provider official rules]. List visible materials, applicable clauses, source dates and uncertainties. Infer no bin type from color, import no foreign rules, operate no hardware and handle no hazardous items.
Evidence limits and cautions
This is a planning demonstration without end-to-end robotic sorting success rates, not completed physical handling.
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
- Do not handle, dismantle or test hazardous waste.
- This exercise controls no robots; use specialist channels for hazardous materials.
Original source: Building the Next Generation of Physical Agents with Gemini Robotics-ER 1.5