Key Takeaways
- Split the weekly report into six prompts: check, compare, explain, summarise, translate, verify.
- Gemini on the web, Microsoft Copilot and Perplexity work in Hong Kong; ChatGPT and Claude do not officially support Hong Kong.
- Never let AI do arithmetic you cannot recheck: ask for formulas and recompute in your spreadsheet.
- Label every reason as confirmed or suspected, so the summary does not invent causes.
- Remove client names and personal data before you upload any export.
A weekly report that takes two hours usually has about ten minutes of real thinking in it. The rest is cleaning an export, working out what moved, and rewriting the same five bullet points in a tone your manager likes.
These six prompts hand that repeatable part to an AI assistant. Copy them in order and swap the [bracketed] parts for your own data.
Which AI tools can turn a spreadsheet into a weekly report in Hong Kong?
Gemini on the web, Microsoft Copilot and Perplexity all accept file uploads and are usable in Hong Kong. ChatGPT and Claude are not on their makers' supported-country lists for Hong Kong, so use them only through an approved company route. The prompts below are written to work in any of these tools.
Order matters for Hong Kong readers. According to StatCounter figures for June 2026, ChatGPT still leads local AI chatbot traffic at 44.9%, followed by Gemini at 23.2%, Perplexity at 15.9% and Copilot at 13.9%. The difference is that only some of these are officially offered here.
| Tool | Hong Kong availability | Where to add the file | Source |
|---|---|---|---|
| Gemini (web) | Open to all Hong Kong users since March 2026; mobile app rolling out later | Prompt box > + > Upload files | unwire.hk |
| Microsoft Copilot | Available; Microsoft's exclusions name mainland China but not Hong Kong | Prompt box > attach (+) | Microsoft Support |
| Perplexity | Available and sold through local telecom bundles | Prompt box > attach; Spaces need Pro | 1010 offer page |
| ChatGPT | Hong Kong not on OpenAI's supported list | Not recommended for work files here | OpenAI Help |
| Claude | Hong Kong not on Anthropic's supported list | Not recommended for work files here | Anthropic |
All six prompts work on free plans for a single weekly file. Paid plans mainly buy larger uploads and saved project spaces.
How do you get AI to check a data export before analysing it?
Ask it to audit the file and report problems without fixing anything. A silent fix is the most common way an AI report goes wrong: duplicated rows, a blank week or a currency column mixed with text quietly changes every total that follows. Read the audit first, fix the source, then continue.
Prompt 1: The data audit
Attached is our [weekly sales / campaign / ticket] export for [date range].
Do NOT analyse or fix anything yet. Only audit the file and report:
1. Number of rows and columns, and the date range actually covered
2. Duplicate rows (show the first 5)
3. Blank or zero values in [key columns]
4. Columns where numbers and text are mixed
5. Values that look impossible (negative quantities, dates in the future)
Return a short table: issue | column | how many rows | example.
Expect output like "14 duplicate rows, all from the Causeway Bay store on Saturday". That one line usually explains a spike before anyone writes a theory about it.
How do you ask AI for week-over-week changes without getting wrong numbers?
Make the AI show its working as spreadsheet formulas, then recompute the key figures yourself. Language models can misread or miscalculate numbers, so treat the AI as the analyst who sets up the comparison and your spreadsheet as the calculator that confirms it. Ask for causes separately, labelled as confirmed or suspected.
Prompt 2: The change table
Using the cleaned file, build a week-over-week table for [metrics, e.g. revenue, orders, conversion rate].
Columns: metric | this week | last week | change | % change.
Flag any metric that moved more than [10]%.
Under the table, write the exact spreadsheet formula you used for each % change, so I can recheck it.
Prompt 3: The driver hunt
For each flagged metric, list up to 3 possible reasons.
For each reason, say which column or rows support it.
Label each reason CONFIRMED (the data shows it directly) or SUSPECTED (plausible but not proven by this file).
Do not invent outside events. If the data cannot explain a change, write: "Not explained by this file."
The CONFIRMED and SUSPECTED labels are the most useful words in this whole pack. They stop a guess about a public holiday turning into a fact in your manager's inbox.
How do you turn the numbers into a summary your manager will actually read?
Give the AI a fixed five-line format and a word limit, then ask for a bilingual version only after the English or Chinese draft is right. A fixed format makes every week comparable, and your manager learns where to look. Translating last keeps the numbers identical across both languages.
Prompt 4: The five-line summary
Write a weekly summary for [manager role] in at most 120 words, using exactly these five lines:
- Headline: the one number that matters most this week
- What moved: 2 changes, with figures from the table
- Why: confirmed reasons first, then suspected ones marked "(to confirm)"
- Risk: one thing to watch next week
- Ask: one decision or resource I need
Tone: plain and direct, no adjectives like "impressive" or "significant".
Prompt 5: The bilingual copy
Rewrite the summary above in formal written Traditional Chinese for [Hong Kong team / regional office].
Keep every number, date and product name exactly as in the English version.
Do not add or drop any point. After the Chinese version, list any term you were unsure how to translate.
Asking the AI to list uncertain terms is quicker than proofreading the whole thing. It points you straight at the two words that need a human.
How do you make AI check its own report for errors?
Run a separate verification prompt that compares every number in the summary against the change table and the original file. Ask for a list of mismatches, not a rewrite. In practice this catches rounding that drifted, a figure copied from the wrong week, and reasons that slipped from suspected to confirmed.
Prompt 6: The pre-send check
Act as a strict reviewer. Compare the summary against the change table and the original file.
List:
1. Any number in the summary that does not appear in the table or file
2. Any reason stated as fact that was labelled SUSPECTED
3. Any metric flagged earlier that the summary leaves out
If everything matches, reply only: "No mismatches found."
For extra safety, paste the summary into a fresh chat with only the original file and run Prompt 6 there. A new conversation has no memory of its own earlier reasoning, so it is less likely to approve its own mistakes.
How do you adapt these prompts to your own weekly report?
Keep the six-step order and change only the bracketed parts: the export type, key columns, threshold and reader. Save the finished set as a reusable note or project instruction, put the data before the instructions in each prompt, and strip client names and personal details before every upload.
- Data first, question last. Anthropic's long-context guidance reports that placing the question after a long document can noticeably improve answer quality. The same habit helps in other assistants.
- Fix the threshold once. If 10% is your team's "worth mentioning" line, keep it every week so the flags stay comparable.
- Remove personal data. Replace customer names and phone numbers with IDs before uploading. Google's Gemini Apps Privacy Hub explains that consumer conversations may be reviewed by people, so personal Gemini accounts are not the place for client records.
Last verified: 17 November 2026.
What should you take from this prompt pack?
A good AI weekly report is a short chain of checked steps, not one clever prompt. Audit first, compare with formulas you can recheck, label every reason, then verify before you send. Run the chain twice and the Monday report becomes a fifteen-minute review instead of a two-hour rebuild.
Empowering AI. Centered on you.
Want the weekly report to build itself?
If you would rather not build and maintain this reporting flow yourself, a UD AI employee can take it on, deployed and maintained for your role. We'll walk you through every step, starting with a free AI employee setup assessment.
Frequently Asked Questions
Is it safe to upload a sales export to an AI chatbot?
Only after you remove client names, phone numbers and other personal data. Hong Kong's PDPO still applies to what you upload, and consumer AI accounts may keep conversations for review.
Which tool handles Excel files best in Hong Kong?
For most teams, Gemini on the web or Microsoft Copilot, because both accept spreadsheet uploads and are offered in Hong Kong. If your company already pays for Microsoft 365 Copilot, Copilot inside Excel keeps the data in your tenant.
Do I need a paid AI plan to use these prompts?
No. A single weekly CSV or Excel file fits within free plans in most cases. Paid plans add bigger uploads, saved project spaces and higher daily limits.
Why not ask for the whole report in one prompt?
One long prompt hides errors inside a confident final answer. Splitting the work lets you see the audit and the change table before anything is summarised, so a bad row is caught early.