[Community]FAQ & Pitfall Guide

Summary: The 85 most common questions across the 53-article series — find answers in one search.

Key Terms You Should Know First

This article is a high-density lookup list: when something breaks, search and you'll find it fast. The first time each term appears below, I'll give you the plain-English version. After that, we treat them as shared vocabulary.

  • macro = a script that runs repetitive steps for you. Usually written in VBA (the built-in language of Excel/Word) or Office Scripts (the newer JavaScript-based Office scripting).
  • PivotTable = a visual table that summarizes raw data by dimension (region / month / category) — no formulas required.
  • DAX = the formula language used in Power BI to write calculation rules, e.g. "compute year-over-year growth."
  • Power Query = the "Data → Get & Transform" entry point in Excel/Power BI, responsible for cleaning messy data.
  • mail merge = a Word feature that sends the same letter/notice to a batch of recipients, automatically swapping in each person's name, amount, or date.
  • M language = the formula language behind Power Query. Looks similar to Excel functions, but it's a different beast.
  • prompt = the text you give an AI to tell it what to do and how to do it. The more specific, the less it wings it.
  • token = the smallest unit a large model bills by. Roughly 1 English word ≈ 1.3 tokens; 1 Chinese character ≈ 1.5–2 tokens.
  • context window = the maximum amount of text an AI can "see" in one go. Longer window = better at long documents.
  • hallucination = when an AI confidently makes things up that sound plausible but are wrong. This is a known LLM trait, not a bug.
  • Copilot = Microsoft's AI assistant baked into the Office suite — available in Word, Excel, PowerPoint, Outlook.
  • AI Agent = an advanced pattern where the AI doesn't just answer — it breaks down a task, calls tools, inspects results, and revises its plan on its own.

I. Pain Points: Why This FAQ Exists

After working through all 53 articles, you're almost certain to hit these "stuck moments":

  • Your boss asks you to "use AI to review this stack of contracts," so you open ChatGPT and ask one question — and get back a wall of unfamiliar English jargon.
  • The Copilot button is nowhere to be found in the top-right of Excel, no matter what you tweak.
  • You write a prompt, and the AI invents a legal clause that doesn't exist.
  • Company IT warns you not to paste customer data into ChatGPT.
  • A VBA macro runs fine on your machine but errors out the moment a colleague opens it.
  • Copilot refuses to show up in Power BI even though your account clearly has a Pro license.

For these scenarios, a Google search won't get you a precise answer, and forum answers contradict each other. This article gives you a "search-and-fix" checklist — broken into four major categories. Each entry is a three-piece kit: keyword → one-sentence answer → necessary action.

II. Goals: What You'll Walk Away With

After reading this article, you'll be able to:

  1. Locate problems in 5 minutes: use the table of contents + Ctrl+F to instantly search through 85 common stumbling blocks;
  2. Resolve 80% of beginner blockers: accounts not responding, buttons missing, macros disabled, prompts going sideways;
  3. Hold the line on data security: know what you can send, what you can't, and how to retract it if you mess up;
  4. Avoid pitfalls as you level up: training custom bots, prompt engineering, and model-upgrade compatibility are all covered.

III. FAQ Master List (Search by Topic)

3.1 AI Tools FAQ (ChatGPT / Claude / Gemini / Grok / Copilot, etc.)

Keyword tags: #login #subscription #no-response #hallucination #lag #model-selection #token #context #network #account

Accounts & Login

Q1: #login ChatGPT keeps spinning or throws "Network Error" on login? Before touching your account: ① try another browser or an Incognito window; ② switch networks (office Wi-Fi ↔ phone hotspot); ③ if it persists, the service is likely under heavy load — try again off-peak.

Q2: #login You see "You've been blocked" or "Access denied"? Stop retrying for a few hours — rapid-fire requests trip anti-abuse protection. If it persists on a normal home connection, contact support; otherwise just use one of your other assistants (Claude, Gemini, Grok) in the meantime.

Q3: #login Can multiple people share one account? Don't do this on the free tier — OpenAI will ban the account outright. The Plus Team Plan has a 2-seat minimum, billed per seat, and each person logs in independently.

Q4: #subscription Is ChatGPT Plus worth subscribing to? At $20/month, it unlocks three things: ① GPT-4o model (free tier only gives you GPT-4o mini); ② file upload and data analysis; ③ DALL·E 3 image generation and the GPTs store. If you only write weekly reports and translate, free tier is enough. If you need to feed PDFs, analyze data, or write code, Plus is essentially mandatory.

Q5: #subscription Plus / Pro / Team / Enterprise — how to choose?

  • Plus ($20/month): for individuals, the most common pick;
  • Pro ($200/month): unlocks o1 pro, unlimited Advanced Voice, and higher message caps;
  • Team ($25/month/seat): 2-seat minimum, billed per seat, for collaboration;
  • Enterprise (quote-based): for corporate procurement — includes SSO, domain lock, data isolation.
  • For a regular office worker, Plus is plenty. Don't jump to Pro just to "try it out."

Q6: #subscription How do I pay for a subscription? Pay directly on the official site with a regular credit/debit card, or via the App Store / Google Play in-app purchase. Never use third-party "top-up agents" or shared accounts — they can leak your chat history or get the account banned.

Q7: #subscription I paid, but the premium features aren't unlocked? Sign out and back in to refresh entitlements; web and App share the same membership, so try the other side; if it still fails, contact support from inside the app with your order number — usually resolved same-day.

Q8: #account After leaving a Team account, can the company see my chat history? Yes — and you can't delete it. Team Plan defaults to Workspace sharing, which means the company admin can theoretically view all conversations. Don't chat about sensitive content on a work account.

No Response / Lag / Errors

Q9: #no-response AI keeps spinning and produces nothing for minutes? Most likely the node is unstable or the request timed out. Three-step diagnose: ① paste the same prompt into Claude/Qwen to see whether the model itself is the issue; ② switch nodes and retry; ③ break the request down (5 sub-questions vs. 50 in one shot).

Q10: #no-response Error: "You've reached the usage limit"? Free users have a per-3-hour message cap (~40 messages). When you hit it, wait or upgrade to Plus. Don't spam refresh — refreshes count as requests too.

Q11: #no-response Error: "Context length exceeded"? The document you fed in is longer than the context window. ChatGPT-4o caps at 128K tokens, roughly 60,000 Chinese characters. Fixes: ① ask the AI to summarize earlier content before continuing; ② split the doc into sections and feed them in batches; ③ switch to Claude (200K tokens) or Gemini (1M–2M tokens).

Q12: #lag Output cut off mid-generation? Two possibilities: ① network jitter — wait a few seconds and click "Regenerate," it will continue from the breakpoint; ② content moderation triggered (your text contains sensitive words) — the AI pretends to generate but is actually halted. In that case, rephrase and retry.

Q13: #lag Is Chinese slower than English? No — model speed difference between Chinese and English is negligible. If you feel it's slow, the bottleneck is almost always your node or too many open browser tabs.

Hallucination & Accuracy

Q14: #hallucination AI invents a paper / regulation / link that doesn't exist? This is the classic LLM trait called hallucination. The AI has no concept of "I don't know" — it will confidently fabricate something plausible. Never directly cite the links, numbers, or cases the AI gives you. Always double-check.

Q15: #hallucination How to reduce hallucinations? Four techniques: ① explicitly tell it "if you don't know, say so — don't make things up"; ② ask it to cite sources; ③ for key facts, have it "outline first, then expand"; ④ replace the pure LLM with Perplexity (an AI search engine with built-in web retrieval).

Q16: #hallucination The Excel formula the AI gave me errors out? The AI's formula-writing leans heavily on training samples, so less common functions (like LET / LAMBDA) get written wrong often. Three-step diagnose: ① review the parameters of each segment; ② press F9 to step through and locate the failure; ③ paste the error back to the AI and have it fix it.

Q17: #hallucination The VBA code the AI gave me errors out? LLMs often assemble VBA function names "from memory." Don't copy-paste and run. Three prerequisites: ① in the VBA editor, run Debug → Compile to check for syntax errors; ② for any method you don't recognize, press F1 to read the docs; ③ paste the full error message back to the AI and have it fix it.

Q18: #hallucination The AI's math answer is wrong? ChatGPT-4o still stumbles on arithmetic, especially multi-step multiplication/division or large numbers. Fixes: ① ask the AI to "show the equation before calculating"; ② run critical numbers in Python (turn on Advanced Data Analysis); ③ double-check with a calculator yourself.

Model Selection & Upgrades

Q19: #model-selection GPT-4o / GPT-4o mini / o1 / o3 — how to choose?

  • GPT-4o: the everyday workhorse — writing, translation, summarization, conversation, all capable;
  • GPT-4o mini: free tier default — fast but downgraded ability, suited for simple tasks;
  • o1 / o3: reasoning specialists — math, code, complex logic; slow and expensive, unnecessary for daily office work;
  • GPT-4.5 (if available): stronger EQ and creative writing, but lower benchmark scores than o1.
  • Mnemonic: 4o for daily, o1 for hard problems, mini for tight budgets.

Q20: #model-selection Which assistant should be my daily driver?

  • Writing / translation / summarization: ChatGPT and Claude are both excellent;
  • Code / data analysis: ChatGPT and Claude trade the crown release by release;
  • Long documents: Gemini leads on context window size (check current official specs);
  • Fast web-connected answers / X integration: Grok.
  • Bottom line: pick one main assistant plus one cross-checker.

Q21: #model-selection ChatGPT vs. Claude vs. Gemini vs. Grok — which is strongest? There's no "strongest" — it depends on the use case (and versions change fast, so treat this as directional):

  • Coding: Claude and ChatGPT trade the lead; Gemini close behind;
  • Long documents: Gemini leads on context window;
  • Creative writing: each has its own strengths — try them yourself;
  • Rule of thumb: start on free tiers, keep at least two models and cross-check answers.

Q22: #model-upgrade The model just upgraded — will my old prompts still work? Yes, but rerun your regression tests. New models usually keep backward compatibility, but occasionally: ① something you disabled in the old prompt is now interpreted the other way; ② default behavior shifts (e.g., more verbose, or more concise). For important prompts, run a regression test after every model upgrade.

Tokens & Billing

Q23: #token Roughly how many tokens is a prompt? Quick estimate: 1 English word ≈ 1.3 tokens; 1 Chinese character ≈ 1.5–2 tokens. A 1,000-character Chinese prompt is roughly 1,500–2,000 tokens.

Q24: #token When the context fills up, does the AI really "forget"? Yes. When context approaches the cap, the AI prioritizes the beginning and the end — the middle gets fuzzy. Fixes: ① manually arrange your prompt so critical info sits at the head and tail; ② turn on "Memory" so the AI retains long-term preferences; ③ feed sections one at a time and ask for a summary between each.

Q25: #token How to save tokens? Four moves: ① cut filler ("please help me" can usually go); ② use abbreviations (HR → Human Resources, but mixing styles hurts clarity); ③ ask for bullet lists instead of long paragraphs; ④ have the AI summarize a long document before you ask questions.

General Usage Tips

Q26: #general How do I get the AI to remember me? Three layers: ① Within a single conversation: stuff background info into the first message; ② Long-term Memory (ChatGPT Memory / Claude Projects): once enabled, the AI remembers your preferences across conversations; ③ Custom GPTs / Projects: bundle instructions + knowledge base into a dedicated bot.

Q27: #general The AI's output formatting is messy. What do I do? Specify the format in the prompt: e.g. "Use Markdown level-3 headings + tables," "Each paragraph under 5 lines." The more specific, the less the AI drifts.

Q28: #general How do I stop the AI from running wild? Add constraints: "Strictly use only the information I provided — don't supplement with external knowledge," "Don't fabricate data; if you're unsure, say so."

3.2 Office Software Integration FAQ (Excel / Word / Outlook / PowerPoint / Power BI / WPS)

Keyword tags: #Excel #Word #Outlook #PowerPoint #PowerBI #WPS #VBA #button #membership #permission

Excel Copilot / AI Features

Q29: #Excel No Copilot button in the top-right corner of Excel? Most likely this isn't your problem — it's a tenant/account configuration issue. Four-step diagnose:

  1. Is your account a Microsoft 365 Enterprise or Family edition? The free personal edition doesn't include Copilot.
  2. Is your Excel version recent enough? You need the Microsoft 365 channel (Monthly Enterprise build 2402 or later).
  3. Where is the file saved? It must be in OneDrive / SharePoint — local files don't trigger the button.
  4. Region settings? Some regions (certain EU countries) restrict Copilot by law — switch to an unrestricted region. If none of this works, contact your company IT to reset the tenant license.

Q30: #Excel Copilot shows up but does nothing when I click it? ① Confirm the network can reach graph.microsoft.com; ② Excel Options → Trust Center → Privacy Settings → change "Diagnostic Data" to "Required"; ③ if a restart of Excel doesn't help, restart your computer.

Q31: #Excel Can Copilot make PivotTables? Yes — and it's one of its strong suits. Just say "make me a sales PivotTable by region and month," and Copilot will build it automatically. But for complex multi-table PivotTables, you'll still need manual work or Power BI.

Q32: #Excel Copilot's formulas keep erroring out? About 80% of Copilot-generated formulas are correct; the other 20% fail because of relative/absolute reference mistakes. Hit Ctrl+Z to undo, then click Copilot's "Try Again" button — the second attempt usually nails it.

Q33: #Excel Want to undo Copilot's changes? Press Ctrl+Z — you can undo up to 100 steps. Or click the "Undo" button in the Copilot sidebar (this undoes the entire Copilot task, not individual cells).

Word Copilot / AI Features

Q34: #Word No Copilot button in Word? Same logic as Excel (Q29) — account/version/location, the three usual suspects. One extra gotcha: Word must be opened on a .docx file. Legacy .doc files aren't supported.

Q35: #Word Copilot rewrote my document into something unrecognizable? Copilot's Rewrite tends to rework large chunks wholesale, not polish locally. Best practice: ① select the paragraph you want changed before invoking Copilot; ② use "Rewrite this paragraph" rather than "Rewrite whole document"; ③ ask the AI for "3 alternative versions" and pick the one you like.

Q36: #Word Does Copilot include citations? Yes — but not by default. Tick the "Include citations" box in the Copilot sidebar to enable them. Mandatory for academic or legal documents.

Outlook Copilot / AI Features

Q37: #Outlook No Copilot in Outlook? Only the new Outlook (One Outlook) supports it. Legacy Outlook (LTSC / 2021 / 2019) doesn't. How to upgrade: flip the "Coming Soon" toggle in Outlook's top-right corner.

Q38: #Outlook Copilot writes emails that don't sound like me? Teach it your voice: ① first click "Match my style"; ② write a few emails yourself first, then let Copilot revise them — the AI learns your common phrasings and signature style.

Q39: #Outlook How do I use mail merge? Stop asking AI to copy-paste every email by hand. Word + Outlook's built-in mail merge is purpose-built for this:

  1. In Word, prepare the template (recipient / salutation / body);
  2. Mailings → Select Recipients → Use an Existing List → pick your Excel contact list;
  3. Preview → Finish & Merge → Send Email Messages. Combine with Copilot: let it draft the body first, then drop it into the template. Your efficiency doubles.

Q40: #Outlook Outlook macros are disabled — how do I enable them? File → Options → Trust Center → Trust Center Settings → Macro Settings → choose "Disable all macros with notification" or "Disable unsigned macros." Strongly recommend against "Enable all macros" — even ones you wrote yourself. A colleague's "invoice.xlsm" could carry malware.

Q41: #Outlook How do I auto-reply to certain emails? Two options: ① Simple rules: Outlook's built-in "Rules and Alerts," triggered by sender / subject / keyword; ② Advanced: VBA's Application_NewMailEx event hook, auto-replying once rules match. Don't let AI write production Outlook rule code on the fly — test it in a sandbox mailbox first.

PowerPoint Copilot / AI Features

Q42: #PowerPoint Can Copilot generate a PPT from scratch? Yes, but quality varies. The right way: ① start with a Word outline (chapter headings + bullet points); ② use Copilot's "Create from Word document"; ③ then manually tune the layout. Ten times better than just saying "make me a quarterly report PPT."

Q43: #PowerPoint Copilot's PPT has no images? Explicitly request images in your prompt: "at least one image per slide," "use icons + flowcharts to illustrate steps." Or generate images with DALL·E and paste them in.

Q44: #PowerPoint Is VBA reliable for PPT automation? Reliable — but only recommended for batch formatting (changing fonts/themes in bulk, adding logos to multiple decks). For PPT content generation, use Copilot + manual tweaks — VBA-generated PPTs tend to look ugly.

Power BI Copilot

Q45: #PowerBI I can't see Copilot in Power BI? Most likely one of these reasons:

  1. License not enough: you need Power BI Premium Capacity (P1 and above) or Premium Per User (PPU); regular Pro accounts don't see it;
  2. Your workspace isn't in Premium Capacity;
  3. The dataset doesn't have Q&A natural-language query enabled;
  4. The tenant admin hasn't turned Copilot on. The most common traps are 1 and 2 — go to Power BI Service → Workspace → check if the capacity is Premium.

Q46: #PowerBI Copilot's generated DAX measures are wrong? Very common. AI-written DAX often ignores filter context. Three-step fix:

  1. Take a screenshot of the AI's DAX and your data model, paste them back, and let the AI revise;
  2. Run EVALUATE in DAX Studio to inspect the returned results;
  3. Use CALCULATE to explicitly declare the filter conditions.

Q47: #PowerBI Can Copilot write Power Query M? Yes, and it's recommended. M language is niche — letting AI generate Table.Group / Table.Pivot / Table.AddColumn for you saves half the time. But: ① break complex logic into steps, don't pile it into one mega-expression; ② add comments at every step so you can revisit later.

Q48: #PowerBI Power BI data refresh fails? Three common causes: ① data source connection broken (database password expired, API token expired); ② gateway offline; ③ data volume exceeds the capacity quota. Troubleshooting order: refresh in Power BI Desktop locally → read the error code → check the official docs → contact IT.

WPS AI / Domestic Office Software

Q49: #WPS What's the relationship between WPS AI membership and WPS Super Membership? Two separate systems — don't buy the wrong one:

  • WPS Super Membership: traditional features (cloud storage, PDF, templates) — does NOT include AI;
  • WPS AI Membership: adds AI capabilities (smart writing, reading, PPT generation) on top of Super;
  • WPS AI Mega Membership: adds multi-device sync and higher AI quota on top of AI. Decision rule: if you need AI to write docs/PPTs, you must buy AI Membership or Mega.

Q50: #WPS WPS AI vs. Microsoft Copilot — which is stronger?

  • Chinese writing / domestic templates: WPS AI wins;
  • Formulas / VBA / cross-software integration: Copilot wins;
  • Long documents / PDF parsing: WPS AI wins (handling Chinese PDFs is WPS's home turf);
  • Enterprise compliance / domain accounts: Copilot wins.
  • Mnemonic: Chinese office work → WPS. Global teams with strict IT → Copilot.

Q51: #WPS How do I disable AI in WPS (worried about misclicks)? WPS Client → Settings → Config Tools → AI Settings → turn off "Smart Assistant." Note: after disabling, some menu items will disappear — that's normal.

Q52: #WPS Is WPS's VBA the same as Excel VBA? ~90% compatible — most code runs as-is. Differences: ① WPS doesn't support system hooks like Application.OnKey; ② WPS has its own WPS.* extension objects (cloud doc-related); ③ WPS is missing some library references (like Microsoft Scripting Runtime) — you have to manually check them.

VBA / Office Scripts General

Q53: #VBA The macro runs on my machine but errors on a colleague's? A classic trap. Causes ranked by probability:

  1. Different macro security levels: your colleague set "Disable all macros";
  2. Missing reference libraries: your machine has a library installed (e.g., Adobe Acrobat) theirs doesn't;
  3. Hard-coded file paths: your C:\Users\you\Desktop\xxx doesn't exist on theirs;
  4. Excel version mismatch: code written on 365 may not run on 2016 (e.g., LET / LAMBDA). Permanent fixes: ① use relative paths; ② stick to built-in Excel objects where possible; ③ when asking AI to rewrite code, add a version check If Application.Version >= 16 ....

Q54: #VBA The AI-written VBA errors with "Compile error: Method or data member not found"? Nine times out of ten, a misspelled property or method name. AI often writes attributes that "look right" but don't actually exist. Fixes: ① in the VBA editor, press F2 to open the Object Browser and search for the method; ② paste the line number from the error back to the AI and have it fix it; ③ when calling a macro via Application.Run, remember to wrap macro names in quotes.

Q55: #VBA Office Scripts vs. VBA — which is better?

  • VBA: the old veteran — broad compatibility, but only on Windows/Mac desktop Office; the web version can't use it;
  • Office Scripts: TypeScript-based, cross-platform (Windows / Mac / Web), but some VBA APIs aren't available (e.g., form controls, SendKeys).
  • Mnemonic: heavy local automation → VBA. Lightweight cross-platform scripts → Office Scripts.

Q56: #VBA How do I debug VBA? Five tools in the kit: ① F8 to step through; ② F9 to set breakpoints; ③ Debug.Print to output to the Immediate window; ④ MsgBox for pop-up variable inspection mid-run; ⑤ let the AI review your code logic (paste the whole thing back and ask it to find bugs).

Q57: #VBA How do I make VBA run faster? Eight tricks: ① turn off ScreenUpdating; ② use arrays instead of cell-by-cell operations; ③ disable auto-calculate with Application.Calculation = xlCalculationManual; ④ batch-write rather than loop-write; ⑤ use Find instead of looping to compare; ⑥ SQL queries on Excel data beat cell loops; ⑦ avoid repeated .Select/.Activate; ⑧ when needed, switch to C# (Excel-DNA / VSTO).

Q58: #VBA Why is my macro mysteriously being modified? Have you used a "universal template" handed out by someone? Many templates embed malicious code (auto-sending emails, auto-downloading scripts). First step after downloading any template: open the VBA editor, review every module, and delete any procedure you don't recognize.

3.3 Data Privacy FAQ (Compliance, Logging, Retraction, Training Data)

Keyword tags: #privacy #compliance #data-security #training #retraction #company-files #customer-data #redaction

Q59: #privacy Can I upload company files to ChatGPT? No. No. No. Important things bear repeating. The only exceptions are:

  • The company has explicitly approved AI tool use (with compliance sign-off);
  • You're using an enterprise edition (ChatGPT Enterprise / M365 Copilot Enterprise / Azure OpenAI), which provides data isolation;
  • The content has been de-identified (names, IDs, customer names, amounts, contract numbers stripped out). A normal ChatGPT account = your company file in plaintext to OpenAI. That violates PIPL (Personal Information Protection Law) and the Data Security Law.

Q60: #privacy How do I know whether my company allows AI use? Three steps:

  1. Check company IT policy: is there an "AI Tool Usage Guideline" or "Generative AI Management Policy"?
  2. Get dual approval from your direct manager and the IT department;
  3. Prefer AI procured by the company (e.g., M365 Copilot Enterprise / an internal LLM platform) over personal accounts. When in doubt, don't use it.

Q61: #privacy Will my ChatGPT conversations be used to train models?

  • Free + Plus: yes by default, unless you turn off "Improve model for everyone" in settings;
  • Enterprise / Team / API: no — data is isolated and not used for training. Path to disable: Settings → Data Controls → turn off "Chat history & training."

Q62: #privacy Is turning off "Improve model for everyone" enough to be safe? Not quite. Your conversations may still be retained by OpenAI for 30 days for abuse monitoring (unless you're on Enterprise / API). For truly sensitive content, use the API (call OpenAI via API key in code) — no conversation record kept.

Q63: #privacy I uploaded something by mistake — how do I retract it?

  • ChatGPT Web: conversation list → click the conversation → trash icon in the top-right → delete. It disappears from your account, but OpenAI's backend may still have backups (legally required for a retention period);
  • API calls: no conversation record by design;
  • M365 Copilot Enterprise: contact the company admin to delete it from the compliance log. If you really did upload a customer list / financial data: ① delete the conversation immediately; ② notify the company compliance department immediately; ③ keep a screenshot of the deletion as evidence of "good-faith remediation."

Q64: #privacy What information is "absolutely forbidden" to send? By danger level:

  • 🔴 Absolute prohibition: customer ID numbers / phone numbers, bank card numbers, passwords, unpublished financial reports, unpublished contracts, internal personnel files, medical records;
  • 🟡 De-identified is okay: customer names (replaced with codes), amounts (generalized), contract terms (specific numbers stripped);
  • 🟢 Safe to send: public information, your own first drafts, de-identified samples. Mnemonic: ask yourself "Would I be fired if this leaked?" — if yes, don't send it.

Q65: #privacy My company bought M365 Copilot — am I in the clear? 80% in the clear. M365 Copilot Enterprise promises:

  • Data is not used for training;
  • Data is stored in a tenant-isolated environment;
  • Compliance logs are auditable. But there's still 20% residual risk: ① an employee with too much access feeds Copilot documents they shouldn't; ② Copilot's summaries leak Person A's document content to Person B (the over-permission problem). Mitigation: IT sets least-privilege access. Copilot can only see what you could already see.

Q66: #privacy Using domestic LLMs (Qwen / Wenxin / Kimi / Doubao) — is that compliant? Depends on the use case:

  • Content not involving personal info / trade secrets: compliant, and you don't need to route data abroad;
  • Customer data / financial data: still requires de-identification, plus company sign-off;
  • Cross-border business: mind the cross-border data compliance — ChatGPT uses overseas servers, which counts as cross-border data transfer. Mnemonic: domestic models solve "do I send data abroad?" — they do not solve "do I need to redact?"

Q67: #privacy Who owns the copyright on AI-generated code / documents? A legal gray area, but there are common understandings:

  • The U.S. Copyright Office ruled in 2023: pure AI-generated content is not copyright-protected;
  • Human-authored and substantially edited portions can be copyright-protected (e.g., "AI draft + I heavily revised");
  • For commercial documents, keep the revision history — proof that you made substantive contributions.

Q68: #privacy Customer names appear in the AI's output — should I worry? Yes. Even if you never uploaded that customer's name, the AI might "know" them from training data and slip the name into a generic template. Always run a "redaction three-check" before sending anything out: ① personal names; ② company names; ③ amounts / numbers.

3.4 Advanced FAQ (Prompt Engineering, Custom Bots, Model Upgrade Compatibility)

Keyword tags: #prompt #GPTs #agent #training #advanced #model-upgrade #fine-tuning

Poor Prompt Results

Q69: #prompt The AI's output is never satisfying — what should I do? 90% of the time, the prompt isn't specific enough. Use the 5-Element Prompt framework:

  1. Role (who you are): "You are a financial analyst with 10 years of experience"
  2. Task (what to do): "Review this contract for risks"
  3. Input (what material): the contract body
  4. Output format (how to present): "Markdown bullets + risk levels"
  5. Constraints (what NOT to do): "Don't invent clauses; don't give legal advice" Missing any one of the five and the AI will "freestyle."

Q70: #prompt Are longer prompts always better? No. Prompts have diminishing returns — the first 200 words matter most; every additional 100 words adds < 5% effectiveness. Sweet spot: 500–1,500 words. Over 3,000 words, split into a "system prompt + multi-turn dialog."

Q71: #prompt How do I get the AI to talk like me? Few-shot examples outperform "act like me" instructions by 10×:

  1. Write 3–5 sample paragraphs in your own style;
  2. Append to your prompt: "Answer in the style of these examples: [Example 1] [Example 2] [Example 3]";
  3. Let the AI generate, then manually refine and add it as a new example — forming a feedback loop.

Q72: #prompt Are Chinese prompts or English prompts better? Chinese is fine in most cases, but:

  • For specialized terms appearing for the first time, English is more precise (asking the AI to "translate it into plain language" just adds a round trip);
  • For code / formulas / database fields, English is mandatory;
  • For international models (GPT / Claude / Gemini), English is slightly better understood. Trick: mix Chinese and English, with key terms in English parentheses — e.g., "Build a Power Query (data transformation tool) to merge CSV files."

Q73: #prompt The AI ignores my format requirements. What do I do? Three moves:

  1. Give a counter-example in the prompt: "Don't use Markdown tables; use level-3 headings + lists instead";
  2. Require it to outline first, then expand;
  3. If the output is messy, just say "Please strictly follow this format and regenerate: [your template]."

Q74: #prompt The AI answers the wrong question? Most likely the question itself is vague. Self-check three questions:

  1. What exactly do I want the AI to do? (What's the verb?)
  2. What information does the AI need to answer? (Did I provide the context?)
  3. How will I judge whether the answer is right? (What's the success criterion?) If you can't answer these three, the AI can't either.

Training Custom Bots (GPTs / Projects / Gems)

Q75: #GPTs How do I train a custom bot for my company? ChatGPT's "Create a GPT" (GPTs) in three steps:

  1. Instructions: write a 200–500 word system prompt defining the bot's role, capabilities, and limits;
  2. Knowledge: upload PDFs/docs as a dedicated knowledge base (up to 20 files);
  3. Capabilities: check the abilities you want enabled (Web browsing / DALL·E / Code Interpreter). When done, save as "Only me" or "Public."

Q76: #GPTs What's the most effective way to use a GPTs knowledge base? Five pitfalls to avoid:

  1. Documents should be well-structured (with TOCs and section headings). Pure scanned PDFs have low recognition rates;
  2. No single file over 50MB;
  3. Update regularly — GPTs don't auto-refresh their knowledge base;
  4. Put the critical info in Instructions; treat the knowledge base as a backup;
  5. When testing, ask for specific details from the knowledge base (e.g., "What's on page 3, section 2?") — if it answers, the bot is working.

Q77: #GPTs Can GPTs call external APIs? Yes, but you need to configure Actions:

  1. Write an OpenAPI specification (the API description file);
  2. Upload it under the GPTs "Actions" tab;
  3. Configure authentication (API key / OAuth). Advanced play: let GPTs query internal company databases, submit tickets, send email reminders. Note: every Action call requires user authorization once — it can't run fully unattended.

Q78: #GPTs Claude Projects vs. ChatGPT GPTs — which is stronger?

  • Claude Projects: knowledge base up to 200MB × N files, supports a custom system prompt per Project;
  • ChatGPT GPTs: 20-file knowledge base cap, supports Actions (external APIs);
  • Gemini Gems: similar to GPTs, knowledge base plugs into Google Drive.
  • Mnemonic: external integrations → GPTs. Big knowledge base → Claude. Google ecosystem → Gems.

AI Agents

Q79: #agent What is an AI Agent? How is it different from a normal chat? Normal chat = you ask, AI answers, one-shot. AI Agent = the AI itself runs a loop of "break down the task → call tools → inspect results → revise the plan." You only state the goal. Example: Normal chat, you say "Help me find the cheapest flight from Beijing to Shanghai tomorrow" — the AI gives advice. With an Agent, you say "I'm going to Shanghai tomorrow" — the Agent checks the weather, looks up flights, looks up hotels, adds it to your calendar, writes a memo, and reminds you.

Q80: #agent How do I build a simple Agent? Minimum-viable stack of three:

  1. Brain: GPT-4o / Claude 3.5;
  2. Tools: let the Agent search the web (Web Search), run code (Code Interpreter), read files;
  3. Loop: each step is "think → do → check → revise." Tool recommendations: Coze (ByteDance, available in China) / Dify (open-source) / n8n / Microsoft Copilot Studio. Don't build from scratch — configuring an existing platform is good enough.

Q81: #agent My Agent is stuck in a loop and won't stop. Help? Almost every Agent hits this wall. Three fixes:

  1. Set a maximum step count (e.g., force stop after 10 steps);
  2. Set an exit condition (e.g., "Stop when the result contains 'done'");
  3. Add a human confirmation node (pause and ask the user before critical steps). Without these three, your Agent is out of control.

Model Upgrades & Compatibility

Q82: #model-upgrade GPT-4o upgraded to GPT-4.5 — do my old prompts still work? Yes, but run regression tests. Behavioral shifts in the new model:

  • It may be more verbose (auto-adding more explanation);
  • It may be more conservative (refusing edge cases the old model would handle);
  • It may be more "politically correct" (rejecting certain prompt-injection tricks). Field test: run 10 historical prompts through the new model, compare outputs. If the divergence exceeds 20%, adjust your prompts.

Q83: #model-upgrade Will my saved GPTs be affected? Yes — model upgrades affect them. When OpenAI upgrades the underlying model, all GPTs automatically switch to the new model — you can't pin a specific underlying model. Workarounds: ① adding "Please behave like GPT-4" in Instructions has limited effect; ② for critical GPTs, fork a copy and test it manually.

Q84: #fine-tuning Can I fine-tune a model on my own data? Yes, but the bar is high:

  • OpenAI Fine-tuning: supports GPT-4o mini / 3.5, needs 50+ sample pairs, training costs range from tens to hundreds of dollars per run;
  • Open-source model fine-tuning (Llama / Qwen / DeepSeek): needs GPU + ML expertise, but cheaper and more controllable;
  • Not recommended for regular users: 99% of scenarios are well-served by RAG (Retrieval-Augmented Generation) + a good prompt — fine-tuning isn't necessary.
  • Mnemonic: try RAG first, fine-tune only if that's not enough, train from scratch only as a last resort.

Q85: #fine-tuning What is RAG? How do I set one up? RAG = Retrieval-Augmented Generation — it lets the AI search your knowledge base before answering. Minimum-viable recipe:

  1. Split your PDFs/Word docs into 500-character chunks;
  2. Convert them to vectors (mathematical semantic representations) using an Embedding model;
  3. Store them in a vector database (Chroma / Milvus / Pinecone);
  4. When the user asks a question, retrieve the 5 most relevant chunks first, then feed them to the AI;
  5. The AI answers based on those 5 chunks. Toolchains: Dify / Coze / FastGPT / n8n all have RAG built in — configure and you're good to go.

IV. Pitfall Guide (Real-World Lessons)

This section is a "lifesaving" list of pitfalls, ordered by how often they happen.

Pitfall 1: Letting AI Generate "Important Outbound Documents" and Sending Them Straight Out

Scenario: Your boss asks you to send a customer notice; you have AI write it and copy-paste the draft straight into the send box. Consequence: The AI invents non-existent policy clauses, a wrong company address, or leaves typos uncorrected → customer complains → you take the fall. How to avoid: AI drafts → you revise → you review → then send. AI is a "first-draft generator," not a "final-publisher."

Pitfall 2: Feeding Unredacted Customer Data to AI

Scenario: You ask AI to help analyze a sales spreadsheet, and you upload it without removing customer phone numbers. Consequence: Violates PIPL; the company can be fined anywhere from tens of thousands to millions. How to avoid: De-identify every field containing personal info before upload (phone → 138****0000; name → Mr. Zhang). When in doubt, don't send it.

Pitfall 3: Trusting AI-Generated VBA Without Debugging, Then Sending It to Colleagues

Scenario: You have AI write a file-operation macro, then send it to 10 colleagues to run. Consequence: AI got the path wrong, deleted files on 5 colleagues' desktops. How to avoid: Always test in a sandbox first → ask AI to add Debug.Print for key variables → read through every line yourself → only then distribute.

Pitfall 4: Using Free ChatGPT to Write Company Contracts

Scenario: Sales asks AI to help draft a contract; they directly upload a customer contract template on the free tier. Consequence: Contract terms, pricing, and customer names all get ingested by OpenAI for training. How to avoid: For company business, use the Enterprise edition / Azure OpenAI / domestic private deployment. Free tier is for non-sensitive content only.

Pitfall 5: Sending VBA Macros Unprotected

Scenario: You wrote a great macro and send it to a colleague. Consequence: The colleague tweaks a couple of words, claims "this is my work," and takes the credit. How to avoid: Add at least basic protection before release (Tools → VBAProject Properties → Protection → set a password). Note: VBA passwords are not absolutely secure — pros can crack them. If you really need confidentiality, convert to .exe or .xll add-ins.

Pitfall 6: Using AI-Generated Excel Formulas Without Verification

Scenario: You ask AI to write a VLOOKUP; it returns =VLOOKUP(A2,Sheet2!A:F,3,FALSE). Consequence: One character missing, wrong sheet name, column index counted from 0… the formula returns #N/A or #REF!, and the whole report is wrong. How to avoid: Three things every formula needs: ① check the reference range; ② press F2 in the cell to see whether the references highlight correctly; ③ use Formula → Evaluate Formula to step through.

Pitfall 7: Skipping Regression Tests After a Model Upgrade

Scenario: GPT-4o upgrades to GPT-4.5; you keep using your old prompts without changes. Consequence: Old batch jobs that used to work now get refused because the new model is "more cautious" — your automation pipeline breaks. How to avoid: Run a full regression test the week of any model upgrade (core prompts + core automations). If issues surface, adjust the prompts or roll back.

Pitfall 8: Antivirus False-Positives Killing Your Macros

Scenario: You send a .xlsm file to a colleague, and the antivirus deletes it on sight, claiming "macro virus." Consequence: Your colleagues can't use your file. How to avoid: ① Compress to .zip before sending (antivirus doesn't scan inside archives); ② ask IT to whitelist your code; ③ switch to Office Scripts or Python scripts (antivirus is far more lenient on those).

Pitfall 9: Citing AI Output as Academic References

Scenario: You write a paper citing "Smith (2020) in Nature" — straight from the AI's mouth. Consequence: The paper is fabricated; reviewers can't find it; your paper gets retracted. How to avoid: Never cite AI output directly. Every reference must be independently verified against the primary source. For academic work, prefer specialized tools like Elicit / Consensus / Semantic Scholar.

Pitfall 10: Assuming AI "Understands" Your Company

Scenario: You tell the AI "write it following our company convention," and the AI immediately produces a draft. Consequence: The AI has no idea what your company convention is; what you get is a generic template. How to avoid: You must spell out the "convention" — "Our email signature is 'Best regards, Zhang San, Tel 138...'", "Our weekly report must contain three sections: Next Week's Plan, Risk Points, Support Needed." AI can't read minds.

Pitfall 11: Asking AI 50 Things in One Go

Scenario: You cram an entire day's worth of work into one prompt. Consequence: The AI misses half of it, the order is scrambled, and key tasks get glossed over. How to avoid: One prompt, one task. Need to solve 50 things? Spin up 50 conversations. Spend 10 minutes decomposing — save 1 hour of rework.

Pitfall 12: Using "Pinyin / Reverse Prompts" to Bypass AI Limits

Scenario: The AI refuses to write "talking points attacking a competitor," so you switch to pinyin "gong ji." Consequence: You trigger even stricter moderation; your account gets flagged. Even if you slip past, the output is terrible. How to avoid: Don't play games. Reframe the positive goal: instead of "attack talking points," write "a product pitch that highlights our differentiated strengths." That's the problem you actually want solved.

V. The Core Principle

All these FAQs orbit one central principle: AI amplifies; it does not replace.

  • It amplifies your capability (faster writing, faster lookup, faster calculation) — but does not amplify your judgment. Whether to trust and use the AI's answer is always your call.
  • It amplifies your efficiency — but does not amplify your security habits. Data still needs redaction; code still needs review.
  • It amplifies your coverage — but does not amplify your expertise. AI-written formulas can't outdo a ten-year finance veteran; AI-written code can't outdo an architect; AI-written contracts can't outdo a lawyer.

The people who truly use AI well = AI's speed × your judgment × your security habits.

VI. Where to Go Next

After working through this article, here are the high-level topics to tackle next:

  • F1 - From User to Sharer: turn your AI pitfalls into articles and give back to the community;
  • F2 - Build Your Personal AI Office Knowledge Base: combine this article + your own 53-article notes + company materials into a local RAG;
  • C6 - AI Agent in Action: upgrade from "ask AI" to "let AI do it";
  • C4 - AI Knowledge Base: how to build enterprise-grade RAG, manage permissions, and prevent data leaks.

Appendix: 30-Second Quick Lookup Table (Recommended Bookmark)

The Problem You're Facing First Step
ChatGPT not responding Try another browser/Incognito → switch network → retry off-peak
Excel has no Copilot Check account version + save to OneDrive
AI hallucinates content Add "if you don't know, say so" to the prompt
VBA errors out F8 step-through + have AI fix it
Don't know how to use mail merge Word → Mailings → Recipients → Use Existing List
Power BI has no Copilot Check whether the license is Premium Capacity
Not sure about sending company data Use the Enterprise edition + de-identify first
Prompts never satisfy you Use the 5-element framework (Role/Task/Input/Format/Constraints)
Want to build a custom bot Pick GPTs / Projects / Coze — any one — and start configuring
Agent out of control Set max steps + exit condition + human confirmation
Errors after model upgrade Run regression tests on historical prompts
WPS vs. Excel choice Chinese office work → WPS; global collaboration → M365

In short: when something breaks, first Ctrl+F this article by keyword. If you can't find it, ask the community. After you get an answer, come back and add to this article — making this FAQ more complete is something we all share.

Next up: F1 - From User to Sharer: Write Your First AI Office Tutorial — turn the pitfalls you've walked into articles, so others don't walk the same road you did.