Editor-selected existing cases and prompts, separate from AI news. Selection dates use Hong Kong time.
2026-10-21 recheck: five existing, untested cases. ChatGPT web/mobile: TW yes, HK/MO no; Free limits. Firefly web: model/credit limits; Teams: paid licence, policy and transcript. Both: HK/TW conditional, MO unknown. GitHub: limited Free; regional features unconfirmed. Local prices and Traditional Chinese output quality unverified. No API/app equivalence. Check permissions and results.
Describe a subject, setting and camera movement in Firefly to draft a clip for your edit, then inspect it for visual errors.
Prompt
Generate a B-roll shot for [video topic/purpose]. Show [subject] in [setting], [action/state]. Use [lighting and atmosphere] with [one camera movement or a static camera]. Composition: [subject position and whether/where to leave room for text added in editing]. Required or excluded elements: [people, objects, text, logos].
Evidence limits and cautions
- Models, credits and settings vary by plan.
- Check for distorted objects. Do not present generated scenes as records of real events.
Original source: Adobe：Generate videos using text prompts
Use an accessible Teams transcript to ask Copilot for draft decisions, owners and deadlines. Leave unstated details blank and confirm them with participants.
Prompt
From this meeting, separate explicit decisions, action items and open questions. For actions, include an owner and deadline only when stated; otherwise write “Needs confirmation.” Provide traceable support and highlight disagreements. Produce a draft only; do not send messages or create tasks.
Evidence limits and cautions
- Availability depends on licensing, policies and transcription settings.
- Leave unstated owners and deadlines unassigned rather than inventing them.
Original source: Microsoft：Catch up on meetings with Copilot in Teams
Supply a function, expected behavior and existing tests in your editor. Ask Copilot for test scenarios first, then generate and run test code.
Prompt
Based on this function’s requirements and the existing test framework, propose normal, boundary and invalid-input cases first. If requirements and implementation conflict, ask rather than treating current behavior as correct. After I confirm, generate tests and explain what each assertion verifies.
Evidence limits and cautions
- Passing tests does not prove correctness or security.
- Do not expose credentials or run unreviewed code in production.
Original source: GitHub Docs：Writing tests with GitHub Copilot
Start ChatGPT Voice with a specific scenario such as ordering food, take turns speaking, then request alternative phrases to practice again.
Prompt
Help me practice [target language] in [conversation scenario]. My level is [level]; play [role]. Use [preferred pace and turn length; suggest and confirm these based on my level if unspecified] and wait for my reply. Let me finish rather than interrupting every minor mistake. After [practice duration or number of turns], explain the most useful improvements in [feedback language], then let me try again. Ask when you cannot hear clearly instead of guessing my meaning.
Evidence limits and cautions
- Speech recognition and transcripts can be inaccurate; avoid treating them as precise pronunciation scores.
- Use a fictional scenario rather than sharing personal details.
Original source: OpenAI：ChatGPT Voice
Redact private details in a desk photo, upload it to ChatGPT and request visible-item observations, uncertainties and a few reversible organizing steps.
Prompt
From this desk photo, list only object categories you can identify confidently and flag unclear areas. Use [organizing goal, total available time and preferred number of steps; ask if unspecified] to propose reversible steps with estimated times. Prioritize essential steps if time is limited; do not add steps to meet a quota. Prefer existing supplies; ask about budget if a purchase seems necessary and do not buy anything. Recommend moving and grouping only, not discarding documents, and do not guess dimensions.
Evidence limits and cautions
- Image interpretation can be wrong. Do not use it to assess electrical or hazardous-item safety.
- Do not discard important documents or assume AI knows who owns an item.
Original source: OpenAI：ChatGPT Image Inputs FAQ