Multimodal basics: images, PDFs, and voice without magic thinking

Goal: Use uploads and voice as inputs with the same honesty rules as text—plus extra care for PII in screenshots and wrong OCR.

Best for: Beginners attaching PDFs, photos, or using voice on mobile.

Uploads and voice are still guesses about pixels and audio. Verify anything you will act on—especially OCR of legal text.

Multimodal means the model can take more than typed text (images, files, sometimes audio). It does not mean the model “sees truth.” Screenshots can be misread; PDFs can be skimmed wrong; voice transcripts invent words. Verify anything you will act on.

Images and voice are inputs—not proof. The model can misread a screenshot the same way it misreads a paragraph. Describe what you see; ask it to list uncertainties.

Why beginners skip this

Uploads and voice feel like magic until a screenshot leaks a phone number or a PDF “summary” invents a date. Multimodal still needs scrub and Unknowns—same honesty as text.

Habit this week (Asia/Shanghai): scrub one image or PDF before upload; mark two Unknowns in the reply and date the pass.

Step 1 — Say what you want from the file

Name the job before the model free-associates. “What’s in this?” invites invention. “Extract the error code and the button labels you can read” keeps it honest.

I uploaded [file/image].
Task: [extract / summarize / answer question]
Only use visible content. If unreadable, say Unreadable—do not guess.
Flag any PII you notice so I can re-upload redacted.

If the model returns a confident paragraph with no uncertainty list, ask again until Unreadable/Unsure lines appear.

Step 2 — Screenshot / PDF drill

I will attach a screenshot of an error dialog (login page, red banner). Ask for: what text you can read, what you are unsure about, and 3 next checks for me—without inventing account settings I did not show.

Step 3 — Voice

Speak Goal/Format/Boundaries out loud. After transcript, read it—fix homophones before you send. Voice is fast; sending from an unread transcript is how “Tuesday” becomes “today.”

See Voice & mobile for commute workflows. Rule of thumb: if you would not send the transcript as email body, do not act on it yet.

Voice note dump of meeting on phone.
Scrub names before paste. Ask for decisions/open/Unclear only. Spot-check one claim.

Save this card

Card: scrub → upload → mark Unknowns in the reply. Voice and images follow the same honesty as text.

1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)

• [ ] Tools/writes paused
• [ ] Refuse-line dated
• [ ] Artifact path saved
• [ ] Next link chosen (or stop)

Before you start

Sit with three real asks from this week. You will label trust/check/stop for “Multimodal basics: images, PDFs, and”—bring the asks, not slogans. Guide: foundations-multimodal-basics. If you only have a blank practice box, stop and fetch the real job first.

Default posture: draft-only until a human confirms sends/writes. Date your first refuse-line (Asia/Shanghai) before granting new tools.

• Goal in one line: ___
• Allowed tools: ___
• Forbidden: ___
• Human approve needed for: ___

How to use this week

This week: upload one image or PDF after scrubbing; mark unknowns in the reply the same way you would for text on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed foundations-multimodal-basics, put 25 minutes on the calendar and redo once.

Close the loop

After Practice now, write one line: what you will refuse next time. That line is the habit.

OCR is a draft

Treat extracted text like a first draft: useful, untrusted, and never the last word on legal or medical wording.

Screenshot of a spreadsheet crop (redacted headers).
Ask what columns are visible vs Unknown. Spot-check one total myself before sharing.

Red flag

Citing image text without spot-check; treating voice transcript as verbatim without a listen-back—label uncertainty.

Pitfalls

Temptation: skip the habit for foundations-multimodal-basics to “save five minutes.”
Job still: Use uploads and voice as inputs with the
Do the smallest correct pass: Done visible · one checkable artifact · Unknowns labeled [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: the artifact + the refuse-line you said aloud.

Practice now

Live card: apply “Multimodal basics: images, PDFs, and voice without magic thinking” to a mistake from this week; write the fix habit (Asia/Shanghai).

Today: before upload, write the extract goal; after, spot-check two facts against the source. Date the check (Asia/Shanghai).

Today card (copy) — foundations-multimodal-basics: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

• [ ] Artifact path written
• [ ] Refuse example would say "yes"
• [ ] Dated Asia/Shanghai

One sentence takeaway

Scrub before upload; mark Unknowns in multimodal replies the same as text. This week: one scrubbed upload with two dated Unknowns (Asia/Shanghai).

Next

Data analysis with files · Voice & mobile · PII habits · When AI sounds sure but wrong

Pair with today: open the first Next link (chatgpt-data-analysis-with-files) only after you dated a refuse-line for this guide (Asia/Shanghai).

多模态基础：图像、PDF 与语音，别神话

目标： 把上传与语音当输入，沿用与文本相同的诚信规则——并额外小心截图里的 PII 与错误 OCR。

适合： 开始附加 PDF、照片，或在手机上用语音的新手。

上传与语音仍是对像素和音频的猜测。凡是你要据此行动的都要核实——尤其是法律文本的 OCR。

多模态 = 模型能接受不止打字文本（图像、文件，有时音频）。这不代表模型「看见真理」。截图可能误读；PDF 可能略读出错；语音转写会造词。凡是你要据此行动的，都要核实。

图像和语音是输入——不是证据。模型会像读错段落一样读错截图。先描述你看见什么；让它列出不确定点。

为什么新手跳过

上传和语音像魔法，直到截图泄露电话、PDF「摘要」编出一个日期。多模态同样要擦洗隐私、标出未知——和纯文本一样诚实。

本周习惯（Asia/Shanghai）：上传前擦洗一张图或一份 PDF；在回复里标出两处未知，并给这次练习写上日期。

步骤 1——说清你要从文件得到什么

我上传了 [文件/图像]。
任务：[提取 / 摘要 / 回答问题]
只用可见内容。若不可读，写 Unreadable——不要猜。
若发现 PII，标出来让我打码重传。

步骤 2——截图 / PDF 演练

我将附上登录页红色横幅的错误对话框截图。请列出：你能读到的文字、不确定之处、给我的 3 个下一步检查——不要编造我没展示的账户设置。

步骤 3——语音

把目标/格式/边界说出来。转写后先读——发出前修正谐音错误。通勤流程见语音/手机指南。

语音流程：我口述一份很乱的周计划。给一条提示词，把语音转写收成 5 条，并把不清楚的时间标成 TBD——绝不编造会议。

保存这张卡

卡片：先擦洗 → 再上传 → 回复里标未知。语音与图片和文本同一套诚实规则。

开始前

打开你真正会用的账号/项目。准备好与「多模态基础：图像、PDF 与语音，别神话」相关的真实材料（不要空白练习框）。

默认姿态：在人确认发送/写入之前只出草稿。授予新工具前，先写一条带日期的拒做句（Asia/Shanghai）。

• 一句话目标：___
• 允许的工具：___
• 禁止：___
• 需要人批准：___

本周怎么用

本周一次多模态：上传前写提取目标，事后对照原文抽查两处。把过/不过写进笔记并注明日期（Asia/Shanghai）。

闭环

「现在就练一次」后写一行：下次你要拒绝什么。那一行才是习惯。

OCR 是草稿

把抽出的文字当第一稿：有用、不可信，绝非法务或医疗措辞的终局。

红旗

未抽查就引用图中文字；把语音转录当逐字稿不听回。误差未标——补标。卡上写不清就不继续——先补检查项再跑。

常见坑

任务：Use uploads and voice as inputs
场景要点：三张真实请求便签在手；练习编号 3/4。
边界：仅草稿；未知 [?]；禁止悄悄发送；禁止编造指标。
返回：可检查产出 + 过/不过理由一行。

任务：Use uploads and voice as inputs
场景要点：三张真实请求便签在手；练习编号 4/4。
边界：仅草稿；未知 [?]；禁止悄悄发送；禁止编造指标。
返回：可检查产出 + 过/不过理由一行。