Turn a 100-page PDF into notes you can actually use
Skills used in this guide (3)
每当用户要对 PDF 文件做任何操作时使用。包括从 PDF 读取或提取文本与表格、合并多个 PDF、拆分 PDF、旋转页面、添加水印、创建新 PDF、填写表单、加密解密、提取图片，以及对扫描件做 OCR 使其可搜索。当用户提及 .pdf 文件或要求生成 PDF 时使用。
npx skills add anthropics/skills -g当用户想创建、读取、编辑或处理 Word 文档（.docx 文件）或 Word 模板（.dotx 文件）时使用。触发词包括：Word 文档、.docx、.dotx，或要求生成带目录、标题、页码、信头等排版的专业文档。也用于从 .docx/.dotx 提取或重组内容、插入或替换图片、查找替换、批注修订，或转换为精排版 Word 文档。若用户要的是报告、备忘录、信函、模板等 .docx 交付物时使用。不要用于 PDF、表格、Google Docs 或与文档生成无关的通用编程任务。
npx skills add anthropics/skills -g引导用户通过结构化流程共同撰写文档。当用户想写文档、提案、技术规格、决策文档或类似结构化内容时使用。该流程帮助用户高效传递上下文、通过迭代打磨内容，并验证文档对读者是否有效。在用户提及写文档、创建提案、起草规格或相关文档任务时触发。
npx skills add anthropics/skills -g
"Read this PDF" is the easiest request to get wrong. Two reasons: the context window will not hold it, and the model will invent the parts it never saw. The reliable order is extract, then target, then ask, with a page number on every claim.
Who this is for
Good fit
- Reports, papers, contracts, prospectuses or tenders over 50 pages
- You have specific questions, not a vague chat
- You need to hand the findings to someone else, so page references matter
Bad fit
- Scanned files without OCR (run OCR first)
- Information that lives in layout (tables spanning pages, formulas, handwritten notes)
- Documents under 10 pages: just paste the whole thing
Steps
1. Install the skill
git clone --depth 1 --filter=blob:none --sparse https://github.com/anthropics/skills
cd skills && git sparse-checkout set skills/pdfCopy skills/pdf/ into your client skills directory.
2. Map the structure first
Analyse report.pdf:
1. Page count, is there a table of contents, is there an extractable text layer
2. List top-level sections with start pages
3. Mark page ranges that contain tables
Report structure only, do not summarise content.If it reports no text layer, run OCR before anything else.
3. Target by question, never read it all
Ask at most three questions per round, require (p.XX) after each claim, and explicitly allow "not mentioned".
4. Produce page-annotated notes
Ask for a three-line conclusion, bullets with page numbers, and an open-questions list.
5. Verify the numbers that matter
Spend three minutes checking amounts, dates and ratios against the original pages.
Traps
| Trap | Symptom | Fix |
|---|---|---|
| Hallucination | Fluent claims that are not in the file | Require page numbers and allow "not mentioned" |
| Empty extraction | Blank or garbled output | OCR first |
| Broken tables | Numbers run together | Ask for one table as CSV |
| Page drift | Citations off by two pages | Ask for both document page and printed page |
| Too many questions | Answers bleed together | Max three per round |
FAQ
Can it read 100 pages at once? Usually no, and it is not the best approach anyway.
What about scanned files? OCR first, then run this workflow.
Chinese PDFs? Yes. Quality depends on embedded font encoding; use OCR if you see mojibake.
Comparing two PDFs? Extract each separately, then compare. Do not paste both at once.
Can I cite the output? Cite only the page-annotated parts and spot-check before publishing. AI summaries are an index, not a source of record.
References
Related Scenes
Write PRDs and product specs with AI
The pain of PRD isn't "wrong format" — it's "goal and acceptance unclear, dev guesses". The point: give background, target user, problem; let AI output "feature list + acceptance + edges", not decide "why build" for you.
Write weekly reports and standups with AI
The pain of weekly reports isn't "can't write" — it's "turning a log into outcomes". The point: give AI the raw material (what you did, what's stuck), have it close into "progress / risk / next", not invent content for you.
Translate and localize docs with AI
Machine translation's scare is "fluent but distorted" — wrong terms, wrong tone, stiff local phrasing. The point: give a glossary and target-reader tone, have it "translate + keep proper nouns", then you verify key sentences — not trust the whole blindly.
Handle your inbox and support messages with AI
The red line for email automation is "draft yes, auto-send careful". "Triage + draft + you hit send" is safest; letting AI blast-send means you own the mistake, not it.
Also useful across scenes
Write technical docs and READMEs with AI
The scare with doc-writing is "AI sounds right but the API name is invented". The discipline: have AI write from real code / real docs, then match each paragraph to source. Don't let it invent APIs from memory.
Tailor Your Resume to a Job Description with AI — Without Inventing Numbers
>-