skills

Agent skills I use daily and have hardened on real work. Each one is a folder with a SKILL.md, usable from Claude Code as /name.

中文说明

deep-read

Makes an agent actually read a long source to the end, and prove it.

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:

1. 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).
2. The agent reads every chunk in order, itself — no subagent summaries, no sampling.
3. Notes are written after each chunk, so progress survives context compaction.
4. The notes open with a coverage statement (file, line count, each chunk's range), and flag speech-to-text errors, contradictions, and promises the source never kept.

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 focus on analysis (claims, evidence, mental models). deep-read focuses on completeness: every line read, with proof.
• 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.

---

中文说明

我日常在用、在真实任务里打磨过的 agent skill。每个 skill 是一个带 SKILL.md 的文件夹，在 Claude Code 里用 /名字 调用。

deep-read：让 agent 真的把长材料读完，并证明读完了

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：

1. 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。
2. agent 自己按顺序读完每一块，不交给子 agent 总结，也不抽样。
3. 每读完一块就写笔记，上下文被压缩也不丢进度。
4. 笔记开头写覆盖声明（哪个文件、多少行、每块的行号范围），并标出转写错字、前后矛盾、以及原文承诺了却没兑现的内容。

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 重在分析框架，deep-read 重在读全，并留下证据。

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。