skills
skills is a public collection of agent skills for Claude Code that I use daily and have hardened on real work. Each one is a folder with a SKILL.md, usable from Claude Code as /name.
A recorded demo from max1874.com/skills, where you can try it yourself.
中文说明
deep-read
Makes an agent read a long source to the end, prove it, and actually understand what it says.
Hand an agent a 3-hour transcript and it will usually read part of it — a single read tops out at ~25k tokens, a paragraph-per-line transcript can't be chunked, the first 2000 lines are the default cap, or it just greps and summarizes — and then hand you a summary that looks complete. deep-read fixes the workflow:
- A script turns the source (SRT/VTT subtitles, text, extracted PDF) into a reading copy with short lines and a chunk plan that fits the read limit, estimated per language (CJK text costs ~4x the tokens of English per character).
- The agent reads every chunk in order, itself — no subagent summaries, no sampling.
- Notes are written after each chunk, so progress survives context compaction.
- Once every line is read, the notes open with what it is saying: the main point, how the parts build it, where it is strong or weak, and what is worth taking away. How hard to fact-check depends on the kind of source: a speech is read for its meaning, a market call for its evidence. The notes end with a coverage line (file, line count, each chunk's range).
- A second agent reads the source from scratch without seeing the notes and writes its own understanding; every disagreement is checked against the source. It catches what the first reader is anchored against, like a host's summary recorded as the guest's view.
On a 3-hour YouTube course (~41k words), this surfaced what a skim would miss: $6.41 transcribed as $641, a community size stated two different ways, and two "I'll explain later"s that never came.
How it differs from similar skills
- full-output-enforcement stops agents from writing less than asked.
deep-read stops them from reading less than given.
- dsh-deepread and the "deep reading" framework skills apply an analysis template (claims, evidence, mental models).
deep-read makes every line read, with proof, the floor, and asks for understanding in the source's own terms on top of it.
- Research skills that delegate reading to a background agent trade completeness for speed.
deep-read makes the opposite trade on purpose.
Install
npx skills add max1874/skills --skill deep-read
Or copy the deep-read/ folder into ~/.claude/skills/ (all projects) or .claude/skills/ (one project). Then use /deep-read, or just ask the agent to read something thoroughly.
Requirements: Python 3 (standard library only). For PDFs, pdftotext from poppler.
FAQ
How do I make Claude Code read a long document or transcript all the way through?
Install deep-read, then run /deep-read or ask it to read everything without skipping. The script builds a short-line reading copy and a chunk plan; the agent reads and takes notes chunk by chunk, and finishes with notes that open with what the source is saying and end with a coverage line listing the lines each chunk covered.
Why does Claude Code only read part of a long file?
A single read errors above ~25k tokens, Read returns the first 2000 lines by default, a paragraph-per-line transcript can't be split by line, or the agent greps and summarizes instead of reading. deep-read's script and rules target each of these.
How do I install a Claude Code skill?
npx skills add max1874/skills --skill deep-read, or copy the skill's folder into ~/.claude/skills/ (all projects) or .claude/skills/ (one project). Then call it with /name.
Does it work with Codex or other agents?
Each skill is a standard SKILL.md folder, and npx skills add can install it for other agents (-a codex). prepare.py runs from any agent with a shell, but the reading steps (offset/limit, pages, the read-size limit) are written for Claude Code's Read tool.
Can it read subtitles and PDFs?
SRT/VTT become timestamped paragraphs. PDFs with a text layer go through pdftotext -layout first; scanned PDFs are read in page order with Read's pages parameter, up to 20 pages at a time.
More: max1874.com/skills
中文说明
skills 是一组公开的 Claude Code agent skill，都是我日常在用、在真实任务里打磨过的。每个 skill 是一个带 SKILL.md 的文件夹，在 Claude Code 里用 /名字 调用。
动图录自 max1874.com/skills，那里可以亲手试。
deep-read：让 agent 真的把长材料读完、证明读完了，并且读懂
把一份 3 小时的转写稿交给 agent，它通常只读了一部分就给出一份看起来完整的总结。原因有几种：单次读取超过约 2.5 万 token 会报错；一段一行的转写稿没法按行切块；默认只读前 2000 行；或者干脆 grep 几下就开始总结。deep-read 把读的流程固定下来：
- 用脚本把原文（SRT/VTT 字幕、纯文本、从 PDF 抽出的文字）折成短行，并按估算的 token 数给出分块计划。中文每个字的 token 成本约是英文字符的 4 倍，脚本会分语言估算。
- agent 自己按顺序读完每一块，不交给子 agent 总结，也不抽样。
- 每读完一块就写笔记，上下文被压缩也不丢进度。
- 全部读完后，笔记开头先写它在说什么：主旨、各部分怎么把它说出来、哪里有力量哪里站不住、值得带走什么。事实核对做多少看材料类型：演讲读的是意思，行情判断查的是依据。笔记末尾一行写明读了哪个文件、多少行、每块的行号范围。
- 另一个 agent 不看笔记，从头独立读一遍，写出它自己的理解；每一处分歧都回原文核对。
实测一门 3 小时的 YouTube 课程（约 4.1 万词），读出了略读发现不了的问题：字幕把 $6.41 识别成 $641，社群规模前后说法不一，两处「后面再解释」最后都没解释。
和同类的区别：full-output-enforcement 防止 AI 写得比要求的少，deep-read 防止 AI 读得比给的少；dsh-deepread 等「深度阅读」类 skill 套分析框架，deep-read 以读全、留证据为底线，在此之上按材料本身读懂它。
安装：npx skills add max1874/skills --skill deep-read，或者把 deep-read/ 文件夹复制到 ~/.claude/skills/。需要 Python 3；处理 PDF 需要 poppler 里的 pdftotext。
常见问题
怎么让 Claude Code 把一份长文档或转写稿完整读完？
装上 deep-read，用 /deep-read，或者直接说「全部读完，别跳」。脚本先生成短行阅读副本和分块计划，agent 逐块读、逐块写笔记，最后的笔记开头写它在说什么，末尾一行写明每块读了哪些行。
为什么 Claude Code 读长文件只读了一部分？
单次读取超过约 2.5 万 token 会报错；Read 默认只返回前 2000 行；一段一行的转写稿没法按行切块；或者 agent grep 几下就开始总结。deep-read 的脚本和规则针对的就是这几种情况。
怎么安装 Claude Code 的 skill？
npx skills add max1874/skills --skill deep-read，或者把 skill 文件夹复制到 ~/.claude/skills/（所有项目）或 .claude/skills/（单个项目），然后用 /名字 调用。
能用在 Codex 或其他 agent 上吗？
每个 skill 是标准的 SKILL.md 文件夹，npx skills add 可以装给其他 agent（-a codex）。prepare.py 在任何能跑命令的 agent 里都能用，但读取步骤（offset/limit、pages、读取上限）是按 Claude Code 的 Read 工具写的。
能读字幕和 PDF 吗？
SRT/VTT 会转成带时间戳的段落。有文字层的 PDF 先用 pdftotext -layout 转成文本；扫描版用 Read 的 pages 参数按页序读，每次最多 20 页。
更多：max1874.com/skills
