system-study

System Study — systematized learning material generation

Tool adaptation

In the text, AskUserQuestion、Agent、Read、Write are capability names; use the equivalent tools on the current platform. When there is no multi-select interface, list the slice options in a single message and ask the user to reply with a combination of numbers; do not ask again about information that is already clear. Sub-agents execute in batches according to the platform's concurrency limit; all inputs and prompts must be fixed first, and later batches must not be injected with the conclusions of earlier batches. When the platform has no sub-agent capability, explain the limitation, then determine with the user how to conduct research from each perspective one by one.

You are about to learn a new field. The AI independently researches + builds a system + produces HTML learning material.

Essential differences from existing skills：

Differences from conclusion-oriented research: conclusion-oriented research is "have confusion → seek a conclusion"; this is "want to learn → seek a system"
vs case-radar: that one provides a scattered collection of cases; this one provides a systematized full picture
vs readable-output: that one processes existing materials; this one digs up materials itself + builds the system itself (in the end it can be called to produce the final draft, but that is not the core)

The differentiating lifeline of this skill: In addition to the parts that textbooks always cover, such as "skeleton + cases + engineering practices",mandatorily produce the two blocks "focus points of controversy + list of blind spots"— this is what ordinary learning materials lack, and it is a goldmine for user research → content/products.

Overview of the 5-stage process

Stage	Name	Number of interventions	Output	Detailed operations
1	Topic convergence	1 time (mandatory)	`01-题目确认.md`	reference/stage-1-topic-narrow.md
2	Research plan	1 time (default)	`02-调研计划.md`	reference/stage-2-research-plan.md
3	Sub-agent parallel research	0 times	`03-调研笔记/A-E.md`	reference/stage-3-subagent-templates.md ⭐
4	System construction + outline confirmation	1 time (default)	`04-体系大纲.md`	reference/stage-4-synthesis.md
5	HTML learning material	0 times	`05-学习材料.html`	reference/stage-5-html-spec.md

Default strict logic: Stages 1/2/4 all ask the user. But the user can say once, "do not disturb me throughout, directly produce the HTML", skipping 2 and 4. Stage 1 can never be omitted— if the topic is wrong, all the later work is wasted.

User-authorized "do not disturb me throughout" fast track

If at stage 1 / 2 / any stage the user says "just go ahead / just give me the result directly / you decide / you judge / do not disturb me throughout" — the main thread immediately switches to the fast track:

Stage 1 still must be asked(use AskUserQuestion to pop up slices) — the topic can never be decided by the AI on the user's behalf
Stage 2 still writes 02-调研计划.md for archiving, butno longer asks the user, and goes directly to stage 3
Stage 4 still writes 04-体系大纲.md for archiving, but no longer asks the user, and goes directly to stage 5
Mark "the user has authorized no disturbance throughout" in PLAN.md— this is an audit trail

Switch condition: the user's wordsexplicitly containAuthorization words such as "directly / don't bother me / decide yourself / judge yourself / you decide". Vague expressions (such as "figure it out") don't count—keep asking according to the default strict process.

Core design principles (5)

These 5 principles are distilled from real-world first launches and run through the entire skill's design philosophy. All stages follow them:

Mandatory checkpoints are not skipped—Stage 1 topic convergence can never be skipped. Users can authorize skipping other checkpoints, but if the topic is wrong, everything is wasted.
Force dissent, forbid fence-sitting—Agent E does not accept "both sides are right". Every issue must be given an X/10 leaning score.
Blind spot list as an independent block—Things not yet agreed upon, still being fought over, or where AIs are just repeating each other go in an independent section, not mixed in with other content.
User research perspective leads—Don't follow the "basics → advanced → expert" textbook route; organize by "what can I use this for". Judgment and perspective are allowed; no need to pretend to be objective.
Write HTML independently, don't call readable-output—Learning materials need sidebar navigation + chapter folding + controversy highlighting + an independent blind spot section; the long-form reading structure is not enough.

5 Sub-agent lineup (A-D flexible + E mandatory)

For detailed prompt templates see reference/stage-3-subagent-templates.md。

Agent	Fixed responsibility	Topic adaptation method
A Skeleton	Official docs + schools / mainstream methods	Specific sources depend on the topic (prompting→Anthropic/OpenAI docs; memory systems→each vendor's memory docs; MCP→protocol docs + implementations)
B Frontier	New forms, new trends, latest discussions	Flexible depending on topic—define in one sentence "what is the latest in this field"
C Real thing	Top-tier cases / real code / original text breakdowns	Must read the original text; the number of cases depends on how many credible sources can be found
D Ecosystem	Tool landscape / engineering / application scenarios	Directly connects to the user's product decisions
E Controversy + blind spots	Mandatory, forced dual-track statement + leaning score + blind spot list	Differentiation lifeline—the core difference from learning materials on the market

After the main thread gets the topic, fine-tune each sub-agent's specific research scope according to the topic situation (not changing the responsibility type, but changing "which sources to check"), then dispatch them in batch.

Workflow (what the main thread needs to do)

Step 0: Create working directory

<主题>学习/                  # 默认目录名，可改
├── PLAN.md                  # 总控
├── 01-题目确认.md
├── 02-调研计划.md
├── 03-调研笔记/
│   ├── A-{骨架切片名}.md
│   ├── B-{前沿切片名}.md
│   ├── C-{真物切片名}.md
│   ├── D-{生态切片名}.md
│   └── E-争议与盲区.md
├── 04-体系大纲.md
└── 05-学习材料.html

Step 1: Topic convergence

Use AskUserQuestion(multiSelect: true) to give the user 4 slice options for her to check
Slice descriptions must be specific about "what is covered", not vague
The user can select all (this prompt learning session selected all 4)
Write 01-题目确认.md to record the topic + learning perspective

Step 2: Research plan

List the sources each sub-agent will check (official docs, communities, case libraries, controversial issues)
Have the user review it; items can be cut or added
When the user says "start", enter Step 3
Write 02-调研计划.md

Step 3: Dispatch 5 Sub-agents in parallel

Each sub-agent uses the current platform's independent sub-agent invocation; below Agent is a capability illustration
Before starting, fix all 5 prompts and dispatch them according to the platform's available concurrency limit; when batching is needed, do not inject the previous batch's results into the next batch
See prompt templates at reference/stage-3-subagent-templates.md
Hard constraints that must be included(see anti-patterns.md）：
- Write to file in segments (Write once every 1-2 sections completed)
- Anti-AI-fluff checklist
- Traceable sources
- Controversies must be dual-track + give a leaning

Step 4: Synthesis + outline confirmation

Read all 5 research notes
Design an 8-chapter structure (intro / skeleton / hands-on / advanced / cases / engineering / controversies / blind spots / appendix)
Write 04-体系大纲.md
Let the user review the coordinate system

Step 5: HTML output

According to stage-5-html-spec.md 's design tokens and component specifications
Single-file HTML, all CSS inlined
Must include: fixed left navigation / 4 types of callouts / case details collapsible / controversies with leaning scores / blind spot checklist
Write it all in one go, do not split across multiple Writes(to prevent being interrupted by intermediate states when truncated)

Key hard constraints (be sure to read anti-patterns.md）

Sub-agents must write to files in segments—they cannot wait to write it all at the end. A socket crash will lose all content.
Agent E is forced to dual-track + give leaning scores—"both sides are right" is not accepted.
Case-type Agents are forced to include 3+ original text excerpts + credibility ratings—secondhand cases summarized by AI are not accepted.
AI-fluff blacklist—skip Chinese blogs with an obviously high density of boilerplate.
The main thread does synthesis, not delegated to sub-agents—sub-agents only produce materials.
PLAN.md as master control—update the progress table on every state change.

Flexible parameters

Parameter	Default	Adjustable
Number of slices	4（A-D）+ 1（E）= 5	3-6 (E must be included)
Intervention level	Once each in stages 1/2/4	User can authorize skipping both 2 and 4
Artifact directory name	`<主题>学习/`	User can change
HTML style	Light warm white	Can be extended to dark / dark purple tones (not yet implemented)
Sub-agent dispatch method	Fixed 5 independent prompts, executed according to the platform's concurrency limit	Do not use the previous batch's conclusions to rewrite the next batch's input

Not applicable (clear boundaries)

"Will X work / why Y / evaluate the feasibility of X" → currently available research tools
"What fun cases are there in the X field / let's see what everyone is playing with" → case-radar(scattered points vs system)
"Organize this pile of materials into HTML" → readable-output(processing existing materials vs digging up materials yourself)
"Write an article / produce a first draft" → writing-assistant
"Make a design mockup / UI prototype" → page-solution-design(full-page solution); specific style modifications are adjusted directly based on the existing interface
"PRD / requirements document" → prd-test-writer

Reference examples

Debut hands-on: Project root directory 提示词学习/(a real case that ran successfully on 2026-05-14)

Topic: Prompt learning (all 4 slices selected)
Actual time spent: ~30 minutes
Output: 5 research notes ~180KB → HTML learning material 120KB (1635 lines)
Key lesson learned: see anti-patterns.md 's "Agent B socket crash" case

For each new topic run in the future, it is recommended to record a note in examples-index.md as a reference for subsequent topics.