read
Read PDF page by page with structured output — key arguments, terms glossary, thesis connections. Use when reading academic papers, books, or articles.
yha9806 · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add yha9806/academic-writing-toolkit --skill read技能说明
name: read description: Read PDF page by page with structured output — key arguments, terms glossary, thesis connections. Use when reading academic papers, books, or articles. allowed-tools: Read, Glob, Grep
/read — Academic PDF Reading Skill
Core Principle
Display first, restrain expansion. Show the page content, provide a terms glossary, give a brief summary and thesis connections. Do not auto-expand, auto-search, or auto-record unless the user explicitly asks. The reader controls the pace and direction.
Trigger Words
This skill activates on: read, next page, continue, skip to p.N, read p.N, /read.
PDF limits
Both are enforced in the dsh app and advisory as a plain Agent Skill. Say which one you are in rather than asserting either.
- Maximum 15 pages per invocation. In the app a wider range is denied
before it runs, with
PAGE_RANGE_EXCEEDED. - Maximum 90 pages per session — a context-health budget, counted in the
app from successful reads folded out of the session log, and denied with
PAGE_BUDGET_EXCEEDED. It counts what went through the harness; reading done outside it is invisible to the count. The fold counts calls, not distinct pages, so re-reading a page you already read spends the budget again — say so before repeating a range. - Outside the app nothing counts for you: the limits are a rule you follow imperfectly, and you should not present the budget as tracked.
- When a request approaches the budget, say so and suggest a new session.
Workflow
- Identify the PDF. If the user provides a path, use it directly. If the user names an author or title, search the project's
literature/directory using Glob to locate the file. - Read the specified page(s). In the dsh app call the
read_pdftool —file_path,first_page,last_page— which is the surface the page guards decide on. As a plain Agent Skill, use the host's own file reader with whatever page selection it offers. Default to the next unread page if the user says "next page" or "continue". - Display structured output following the format below.
- Wait for user instruction. Do not proceed to the next page, take notes, or search for related material unless explicitly asked.
Output Format
For each page read, produce the following:
## p.{N} -- {Topic Summary}
### Content
**{Paragraph topic}**
> "{direct quote}" (p.{N})
- {key point}
- {key point}
**{Next paragraph topic}**
> "{direct quote}" (p.{N})
- {key point}
### Key Terms
| Term | Translation | Context |
|------|-------------|---------|
| {term} | {if non-English, provide translation} | {how it is used on this page} |
### Summary
{3-5 sentence summary of the page content. Focus on the argument structure and evidence presented.}
### Connections
- **Thesis link**: {one sentence connecting to the user's research}
- **Previous page**: {one sentence on continuity with previous page, or "First page" if N=1}
- **Open question**: {if any unresolved point or tension is raised}
---
Next page? Take notes? Expand a connection?
Ceremony Control
Emit the ### Key Terms table only when the page introduces genuinely new
terms, and ### Connections only when there is a real thesis link or open
question — skip empty sections rather than filling them for form's sake. For
a page range, one Key Terms table and one Connections block for the range is
usually right.
Handling Multiple Pages
When reading a range (e.g., "read p.10-15"):
- Produce the structured output for each page individually.
- At the end, add a range summary section:
## Pages {start}--{end} Summary {5-8 sentence summary of the range as a whole}
Constraints
- Never auto-record notes. The user must explicitly say "take notes", "record this", or invoke
/note. - Never auto-search for related literature, web sources, or definitions unless the user asks.
- Never launch agents or parallel tasks.
- All extensions require explicit user request -- expanding a connection, comparing with another source, verifying a claim.
- Literature directory is determined from project configuration. Do not hardcode paths.
- Keep connections to 1-2 sentences max. Do not elaborate. The user will ask if they want more.
- Preserve original language for key terms. If the source is in a non-English language, show the original term alongside the translation.
- Quote accurately. Use the exact text from the PDF. If a quote spans pages, note both page numbers.