deep-read

Deep Read

Deep-read non-technical material and output a complete research article. Four-stage pipeline: Capture → Read → Synthesize → Compose.

What you are not

Not a general-purpose URL scraping tool (that is /read skill）
Not a technical source-code analysis tool (that is source-dive）
Does not output half-finished work (always runs the full pipeline and produces a complete research article)

What you are

A four-stage pipeline:Capture → Read → Synthesize → Compose. Input one or more non-technical materials (URL/PDF/file/pasted text), output a complete, self-contained research article.

Trigger conditions

The material type determines the trigger:

Papers (arxiv, journals) → trigger
Book chapters → trigger
News reports, long-form articles → trigger
Interview transcripts → trigger
Industry reports, policy documents → trigger
User shares a PDF/URL and asks for analysis → trigger (when the material is non-technical)

Boundary cases

User says	Trigger?	Reason
"Analyze React's architecture"	No → source-dive	Technical project analysis
"What is this paper about"	No → chat or /read	Only an overview is wanted, not in-depth analysis; if the user adds an intent to analyze, then trigger
"Help me research the implementation principles of transformer"	No → source-dive	Technical implementation, even though "paper" is mentioned
"Read these education reform papers and write a review"	Yes	Non-technical + multiple materials + explicitly wants an article
"Help me analyze this arxiv link"	Yes	Paper URL + analysis request
"Help me take a look at this paper"	Yes	Paper material + implied intent to analyze
"Turn this interview into an article"	Yes	Interview → article
"Help me take a look at this" + shares a PDF	Yes (when the PDF is non-technical)	Sharing a PDF signals intent to analyze

Input parameters

Parameter	Required	Description
sources	Yes	One or more materials (URL/PDF/file/pasted text)
topic	No	Research topic name; if not provided, infer it from the materials
output_dir	No	Output directory, defaults to the current working directory

Output specification

File naming: {主题}-deep-read_{YYYY-MM-DD}.md

Start of fileMust include YAML front matter:

---
title: {主题}
date: {YYYY-MM-DD}
tags: [deep-read, {主题标签}]
sources:
  - {素材来源1}
  - {素材来源2}
status: draft
---

Always run the full pipeline: Capture → Read → Synthesize → Compose. No partial modes are provided. If the user only wants to scrape, use the /read skill; if they only want to read closely without synthesis, discuss directly in the conversation.

Four-stage process

Phase 1: Capture       Phase 2: Read          Phase 3: Synthesize     Phase 4: Compose
┌──────────────┐      ┌──────────────┐       ┌──────────────┐       ┌──────────────┐
│ 保存原始素材   │      │ 精读每个素材   │       │ 综合分析      │       │ 确认锚点      │
│ 识别类型/结构  │ ──→ │ 产出结构化笔记 │ ──→   │ 多视角压力测试 │ ──→  │ 规划章节      │
│              │      │              │       │              │       │ 逐章写作      │
└──────────────┘      └──────────────┘       └──────────────┘       │ 去AI味+博导审稿│
                                                                     └──────────────┘

See the workflow for each stage for detailed steps:

Phase 1: 01-capture.md — Save the original material and identify its structure
Phase 2: 02-read.md — Read closely and produce a Source Brief
Phase 3: 03-synthesize.md — Comprehensive analysis, multi-perspective stress testing
Phase 4: 04-compose.md — Write a complete research article

Key Constraints

No material, no start: When the user provides no material (URL/PDF/file/text), stop and ask the user to provide material. Do not search for the topic on your own, and do not write the article using general knowledge. This is the pipeline's entry gate, not an optional suggestion. See the 01-capture.md Gate section for details.
Always run the full pipeline: No quick/standard toggle is provided. The value of deep-read lies in the full-pipeline output.
Material type boundary: Code/technical projects → source-dive; non-technical material → deep-read.
Native Chinese: The final output must be natural Chinese writing, with no AI flavor.
Self-contained: The article must be self-contained; readers should be able to understand the core arguments without consulting additional material.
Evidence-based: Every claim must be supported by evidence from the material.

Dependencies and Tools

Tool	Purpose	Phase
mcp__web-reader__webReader	Fetch URL content	Phase 1
WebSearch	Search for supplementary material	Phase 1
Read	Read local files and PDFs	Phase 1-2
Agent (background)	Deep-read multiple materials in parallel	Phase 2
Write	Output the article	Phase 4

Gotchas

Problem	Rule
Material URL returns 404 or is behind a paywall	Try the web cache; if it cannot be retrieved, tell the user and do not fabricate content
Material is not in Chinese	Write the Source Brief in Chinese, preserving key terms from the original text; write Compose in Chinese
Material is too long and exceeds the context	Use a background agent to deep-read it in segments, produce a Source Brief, and then write from the Source Brief
The user says "help me analyze this" but provides no material	Stop and ask the user to provide material; do not search on your own and do not write out of thin air (see Phase 1 Gate)
The user only wants to understand the content and does not want an article	This is not a deep-read scenario; just discuss it directly in the conversation