System Study — Systematic Learning Material Generation

• vs long-research: That one is "have confusion → seek conclusions"; this one is "want to learn → seek a systematic framework"
• vs case-radar: That one provides scattered case collections; this one delivers a systematic full picture
• vs readable-output: That one processes existing materials; this one independently gathers materials + builds a system (it can call readable-output for final drafting, but that's not the core)

Overview of 5-Stage Process

01-题目确认.md

02-调研计划.md

03-调研笔记/A-E.md

04-体系大纲.md

05-学习材料.html

Fast Track with User Authorization of "No Disturbance Throughout"

1. Stage 1 still requires consultation (use AskUserQuestion to send options) — the topic can never be decided by AI on behalf of the user
2. Stage 2 still writes

   02-调研计划.md

   for archiving, but no longer consults the user, directly proceeds to Stage 3
3. Stage 4 still writes

   04-体系大纲.md

   for archiving, but no longer consults the user, directly proceeds to Stage 5
4. Mark "User authorized no disturbance throughout" in PLAN.md — this is an audit trail

Core Design Principles (5 Rules)

1. Mandatory checkpoints cannot be skipped — Stage 1 topic narrowing can never be skipped. Other checkpoints can be skipped with user authorization, but a wrong topic makes all work useless.
2. Force dissenting opinions, avoid vague compromises — Agent E does not accept "both sides are right". Each issue must be given an X/10 inclination score.
3. Blind Spot List is an independent section — Content without consensus, still controversial, or repeatedly paraphrased by AI must be in an independent paragraph, not mixed with other content.
4. User research perspective-led — Do not follow the "basic → intermediate → advanced" textbook route; organize content around "what I can do with this". Judgments and perspectives are allowed, no need to pretend to be objective.
5. HTML is written independently, do not call readable-output — Learning materials require sidebar navigation + chapter folding + controversy highlighting + independent blind spot section, which cannot be satisfied by long-text reading structures.

5 Sub-agent Lineup (A-D Flexible + E Mandatory)

Workflow (What the Main Thread Needs to Do)

Step 0: Create Working Directory

<Topic>Learning/                  # Default directory name, can be modified
├── PLAN.md                  # Master control
├── 01-题目确认.md
├── 02-调研计划.md
├── 03-调研笔记/
│   ├── A-{Framework Slice Name}.md
│   ├── B-{Cutting-edge Slice Name}.md
│   ├── C-{Authentic Case Slice Name}.md
│   ├── D-{Ecosystem Slice Name}.md
│   └── E-争议与盲区.md
├── 04-体系大纲.md
└── 05-学习材料.html

Step 1: Topic Narrowing

• Use

  AskUserQuestion

  (multiSelect: true) to provide the user with 4 slice options to check
• Slice descriptions should be specific to "what is covered", not vague
• The user can select all 4 (as in this prompt learning case)
• Write

  01-题目确认.md

  to record the topic + learning perspective

Step 2: Research Plan

• List the sources each sub-agent needs to check (official documents, communities, case libraries, controversial topics)
• Let the user review, they can remove or add items
• Proceed to Step 3 when the user says "Start"
• Write

  02-调研计划.md

Step 3: Dispatch 5 Sub-agents in Parallel

• Each sub-agent is called via an independent Agent (subagent_type: general-purpose)
• Dispatch all 5 at once in the same message (parallel execution)
• Prompt templates can be found in reference/stage-3-subagent-templates.md
• Mandatory hard constraints (see anti-patterns.md):
  • Write to files in segments (Write once after completing 1-2 sections)
  • Anti-AI fluff list
  • Traceable sources
  • Controversies must have dual-track presentation + inclination score

Step 4: Synthesis + Outline Confirmation

• Read all 5 research notes
• Design an 8-chapter structure (Introduction / Framework / Practical Application / Advanced / Cases / Engineering / Controversies / Blind Spots / Appendix)
• Write

  04-体系大纲.md

• Let the user review the structure

Step 5: HTML Output

• Follow the design tokens and component specifications in stage-5-html-spec.md
• Single-file HTML with all CSS embedded
• Must include: fixed left sidebar navigation / 4 types of callouts / collapsible case details / controversies with inclination scores / blind spot list
• Write in one go, do not split into multiple Write operations (to avoid interruption by intermediate states during truncation)

Key Hard Constraints (Be Sure to Read anti-patterns.md)

1. Sub-agents must write to files in segments — Do not wait until the end to write all at once. Socket crashes will cause loss of all content.
2. Agent E must have dual-track presentation + inclination score — "Both sides are right" is not acceptable.
3. Case-focused agents must include 3+ original text fragments + credibility ratings — Second-hand cases summarized by AI are not acceptable.
4. AI fluff blacklist — Skip Chinese blogs with high density of obvious clichés.
5. The main thread handles synthesis, do not delegate to sub-agents — Sub-agents only produce materials.
6. PLAN.md as master control — Update the progress chart every time the status changes.

Flexible Parameters

<Topic>Learning/

Not Applicable (Clear Boundaries)

• "Is X feasible / Why Y / Evaluate X's feasibility" →

  long-research

• "What interesting cases are there in X / See what everyone is experimenting with" →

  case-radar

  (scattered vs systematic)
• "Organize these materials into HTML" →

  readable-output

  (processing existing materials vs independent material gathering)
• "Write an article / produce a draft" →

  writing-assistant

• "Create design drafts / UI prototypes" →

  design-exploration

  /

  macos-product-design

• "PRD / requirement documents" →

  prd-doc-writer

Reference Example

提示词学习/

• Topic: Prompt Learning (all 4 slices selected)
• Actual time taken: ~30 minutes
• Deliverables: 5 research notes ~180KB → HTML learning materials 120KB (1635 lines)
• Checkpoint lessons: See the "Agent B socket crash" case in anti-patterns.md

examples-index.md