learning-pack

Learning Pack · Learning Pack Generator

Turn "I want to systematically learn a topic" into a complete learning pack you can follow step by step, all at once. Give it a topic (such as Loop Engineer, Prompt Engineering, editing, coffee brewing), and it produceslearning path + systematic materials + curated video list + study notes, all saved to disk in ~/Downloads/学习包/。

Positioning and boundaries (key)

Born from deep-researcher, it inherits its bottom lines of parallel research, real sources, no fabrication, and not overwriting subfolders, butthe direction is completely different: deep-researcher is an intelligence engine for Azhen to make topic decisions when writing; this skill is a learning tool for "help me understand a topic clearly".
Hand-off scenarios:
- Writing a paper / literature review / needing GB/T 7714 or APA citations / needing to submit Word → student-researcher
- Topic research / market landscape / competitor analysis / what's been happening with this lately → deep-researcher
- Turning the learning pack into a WeChat Official Account / Xiaohongshu article → irene-content-engine
- Only wanting video transcripts → video-transcriber / youtube-transcript-local
If any of the above needs are hit, hand off directly; do not force it.
All four deliverables are produced by this skill itself, without outsourcing to any downstream skill. Study notes are a built-in capability of this skill, generated directly, without calling ebook-maker.

Configuration files

File	Responsibility
`SKILL.md`(this file)	Stage flow, parameter negotiation, delivery standards, iron rules
`references/学习包模板.md`	Complete templates for the 4 deliverables (00 mind map HTML template, 01/02/03 Markdown)
`scripts/html2png.sh`	Screenshot the mind map HTML into a high-resolution PNG (used for 00)
`scripts/md2pdf.sh`	Convert the MD of 01/02/03 into ebook-layout PDF (hyperlinks preserved, includes Mermaid rendering)
`references/派单模板.md`	Prompt templates for the three parallel Agents in Stage 2

Stage 1: Requirement confirmation (ask all preferences at once before starting)

All points that require the user to decide must be asked all at once before dispatching any Agent or starting generation. For the first 5 items, if they can be inferred from instructions/memory, use the defaults, but learning pack difficulty must appear in the confirmation items; the last 3 items (video platform, note detail level, whether to convert to PDF) must be clearly presented to the user for a decision:

Got it, preparing to make a learning pack for "{topic}". A few things to confirm before starting (use defaults if unchanged):

Learning pack difficulty: Beginner / Intermediate / Advanced (default: Beginner, friendly to ordinary people)

How far to learn: Just for fun as a beginner / Able to get hands-on and do things / Systematic mastery (default: Able to get hands-on and do things)

Your foundation: Complete beginner / A little / Already have relevant experience (default: Complete beginner)

Daily time you can invest: Fragmented 20 minutes / Half an hour to 1 hour / Large blocks of time (default: Half an hour to 1 hour)

Chinese and English materials: Chinese only / both Chinese and English / prioritize first-hand English (default: both Chinese and English, Chinese first)

Which platform for videos(you must decide): prioritize YouTube / both YouTube and Bilibili / Bilibili only

Study note detail level: moderate and thickened (default) / detailed tutorial version

Whether to convert 01/02/03 to PDF(with e-book layout, clickable hyperlinks): no / yes

Reply to me and I'll start.Produce 4 items: study path (mind map diagram) + systematic materials + high-quality video list + study notes, saved to ~/Downloads/学习包/YYYYMMDD-[主题]/。

Difficulty control + three options that must be clearly presented

Study pack difficulty (must be listed in the Stage 1 confirmation): defaultbeginner level. Difficulty controls "how to explain, how deep to filter, and how dense the terminology is", not the learning objectives themselves.
- Beginner level: the first version for ordinary people, with fewer terms, explaining intuition and scenarios first, introducing only a small number of new concepts each time; materials prioritize Chinese/video/official beginner content, and 03 explains each chapter in plain language first before adding technical terms.
- Intermediate: suitable for people who already have some foundation, retaining core terminology and methodological frameworks, adding a small amount of first-hand materials, practice projects, and comparative analysis.
- Advanced: suitable for people with relevant experience, may include official in-depth documentation, papers, source code, complex cases, and industry discussions, but still must be explained clearly by stage.
Video platform (must be asked, cannot decide on the user's behalf by default): prioritize YouTube / both YT+Bilibili / Bilibili only. This directly determines the search scope of the Stage 2 video Agent. If the user has not said, you must ask; only after asking can you dispatch the video Agent.
Note detail level: defaultmoderate and thickened(each chapter has 2 to 3 concept subsections each with examples + one scenario demonstration + 1 to 2 common pitfalls + summary); if the user wantsdetailed tutorial versionthen upgrade (each chapter split into more subsections, concepts explained thoroughly, with runnable code walkthroughs, common error troubleshooting, and end-of-chapter mini exercises, about 1.5 to 2 times the length of the moderate version). Neither level should include illustrations.
Whether to convert to PDF: ask clearly before starting. If the user wants it → automatically convert with md2pdf.sh at the end of Stage 5; if the user says no → do not convert.

Anything already explicitly stated in the user's instructions (such as "beginner level, Bilibili only, want the detailed version, finally convert to PDF") is considered answered and should not be asked again. If the user has not specified difficulty, followbeginner levelto execute.

Quick mode (must be explicitly announced)

When the user information is already sufficient, you may skip the item-by-item Q&A, **but before dispatching the Agent you must send an "inference receipt"** to give a 1-second interruption window;the receipt must list the study pack difficulty, and when the user has not explicitly stated the three items of video platform / note detail level / whether to convert to PDF, you must not fill in defaults on your own and start; you must ask first：

**Stage 1 确认**
- 主题：xxx
- 学习包难度：入门级 / 中级 / 进阶级（默认入门级）
- 目标程度 / 基础 / 每天投入 / 中英文：xxx（可推断默认）
- 视频平台：xxx（必问）
- 笔记详细度：适中加厚 / 详细教程版
- 是否转 PDF：不用 / 要
- 产物路径：~/Downloads/学习包/YYYYMMDD-[主题]/

Dispatching an Agent without a receipt = violation; the receipt omitting the study pack difficulty = violation; starting without clarifying video platform/note detail level/PDF = violation.

Stage 2: Three-way parallel research (mandatory ≥ 3 Agents)

The soul of the study pack is "arranging a reliable study path based on real materials", so information collection must be genuinely researched in parallel, not made up from memory.

Question-decomposition skeleton (3 tracks by default, can be expanded to 5)

Agent 1：系统资料 / 知识体系
  权威入门资料、官方文档 / 教程、经典书目、高质量长文与博客、术语与核心概念。
  目标：搞清这个主题"由哪些模块组成、从哪学到哪、有哪些必懂概念"。

Agent 2：优质视频 / 创作者
  按 Stage 1 拿到的「视频平台偏好」决定搜索范围：
    · 优先 YouTube → 主搜 YouTube，B站 仅补缺
    · YouTube 和 Bilibili 都要 → 两平台都认真覆盖
    · 只要 Bilibili → 只搜 B站，不收 YouTube
  每个必带：视频标题 + 链接 + 博主名 + 粉丝量/订阅量 + 该视频互动量（赞/播放/评论）+ 一句"为什么值得关注"。
  目标：给出按"先看谁、再看谁"排序的视频学习线。

Agent 3：学习路径 / 过来人经验
  社区里"怎么入门 XX""学习路线""新手避坑"的真实讨论（Reddit / 知乎 / V2EX / 即刻 / 论坛）。
  目标：真实的阶段划分、常见卡点、不值得花时间的弯路、怎么验证学会了。

Agent 4（可选）：实战项目 / 练手清单
  这个主题适合拿什么小项目 / 练习题 / 数据集 / 工具来上手。

Agent 5（可选）：用户特别在意的角度
  比如只想要免费资源、只想中文、只想最快上手接活。

For the full Agent prompt template, see references/派单模板.md。

Parallel invocation

Multiple tools must be sent in parallel within the same message. Each Agent prompt must fully carry: topic, Angle, learning-package difficulty, target level and prior knowledge (together they affect how deep the materials go), search terms, must-cover questions, Return format.

Source requirements (inherit the deep-researcher baseline)

At least 1 independent, real source for each of the three buckets: materials / videos / paths; key judgments (e.g. "this is the top pick for beginners") should be corroborated by ≥ 2 sources where possible.
Video links must be real, accessible links, and the blogger's follower count / engagement should be filled in using the real data the Agent finds; if it can't be found, write "not obtained",never fabricate。
Do not treat a single anonymous post as an authoritative path; only mark something as the "widely recognized top pick for beginners" when multiple people recommend it.

Stage 3: Consolidate, sequence the path, identify sequels

Merge Agent results

Archive the raw findings into the three buckets "materials / videos / paths", and deduplicate (same resource from multiple sources → merge).
First, use the Stage 1learning-package difficultyto filter out unsuitable depth: for beginner level, keep only the concepts and resources a learner would use in the first stage; for intermediate level, keep the method frameworks and representative cases; for advanced level, then include deep docs, papers, source code, or complex controversies.
Sequencing the learning path is the core action of this stage: combine Agent 1's knowledge modules + Agent 3's insider stage breakdown to lay out the stage line of "first week / first month / advanced"; then attach Agent 2's videosto the corresponding stages, so the video list and the path align and don't turn into a scattered pile of links.
If resources conflict (A says learn this first, B says learn that first) → in the learning path, mark it as "two mainstream routes", present both, and do not make the hard choice for the user.

Sequel identification

Before Stage 3, check ~/Downloads/学习包/ whether a subfolder with the same topic already exists (prefix match):

Yes → read 01/02/03 in the old directory and include them in the comparison; add -v2 / -v3to the new directory, and add a section to the new learning notes on "what was added compared with last time".
No → skip.

Stage 4: Write the learning package to disk (4 files by default)

Deliverable structure (arranged in the order the learner uses them)

~/Downloads/学习包/YYYYMMDD-[主题]/
  ├── 00-学习路径.png      管「怎么走」：一页式思维导图图片（根节点 + 分阶段分支），看一眼就懂全局
  ├── 01-系统资料.md       管「用什么料」：权威资料清单(标题+URL+推荐理由) + 术语速查表，索引用
  ├── 02-优质视频清单.md   管「看什么视频」：按学习路径排的视频，带博主数据/关注理由/链接
  └── 03-学习笔记.md       管「知识本身」：电子书式正文，把核心知识用例子讲透，阅读用

00 isa one-page PNG mind-map image(generated with HTML layout + Chrome screenshot); 01/02/03 are Markdown and can be converted to PDF on demand in Stage 5 (using ebook layout).

For the full template, see references/学习包模板.md, do not repeat the template in SKILL.md.

Division of labor among the four deliverables (non-overlapping, this is the core constraint of this skill)

Each of the four has a unique bias,the same piece of content appears only once in its home section, don't repeat it across multiple files:

Document	Unique bias	Only put	Strictly forbidden to put (that is someone else's home)
00 Learning path	How to proceed (navigation, one-page mind-map image)	Stage breakdown, order, concept keywords for each stage + passing criteria (made into a one-page PNG mind map)	❌ Do not expand into concept explanations (→03) ❌ Do not pile up resource URLs (→01) ❌ Do not write long passages of text
01 Systematic materials	What materials to use (resource library/index)	One-sentence breakdown of knowledge modules, authoritative resource list for each module (title + URL + type + reason for recommendation), glossary quick-reference table (one-line definition per entry)	❌ Do not explain concepts thoroughly (→03) ❌ Do not lay out timelines/stage planning (→00)
02 Curated Video List	What videos to watch	Video catalog (fields see below)	❌ Do not repeat text materials/concepts
03 Study Notes	The knowledge itself (explained thoroughly)	Ebook-style main text, using plain language + examples to explain each core piece of knowledge clearly, quoting authoritative original words at key points (bilingual)	❌ Do not list a resource library (→01) ❌ Do not do route planning (→00)

Deduplication mantra: In-depth concept explanations → only in 03; one-line glossary quick reference → only in 01; resource URL library → only in 01; stage order and pacing → only in 00. When a concept appears in 00, write only the navigation-level hint "at this stage you need to understand X", without expanding; when a resource appears in 03, attach only that one URL where the original words are quoted, and do not start a separate resource list.

Hard requirements for each of the four deliverables

00-Learning Path.png：One-page mind map image(Irene explicitly wants: not plain text, not the ugly default Mermaid boxes, but a good-looking one-page mind map image). The method islay out a mind map yourself using HTML + inline CSS, then use scripts/html2png.sh to capture it as a high-definition PNG, not writing Markdown, not calling Mermaid.
- Layout:root node on the left (topic) + curves connecting to staged cards on the right(horizontal mind map). Each stage card contains: stage name + duration + concept chips to understand at this stage (about 3 keywords) + a one-line "pass" criterion. Stages are color-coded (a coordinated palette such as tan/cyan/amber/rose), the root node has a tan gradient, and the overall look follows the ebook tan primary color scheme.
- Use SVG to draw Bezier curve connections from the root to each card (first set the canvas size, then set coordinates), canvas about 1180×1040,html2png.sh by default outputs a 2x high-definition image at this size.
- Only include navigation information: learning pack difficulty + stage names + concept keywords + pass criteria.Do not expand concept explanations(see 03),do not spread out resource URLs(see 01). Key points: fit on one page, not crowded, good-looking > complete information. For beginner level, keep concept chips per stage to 2 to 3, to avoid filling a screen with terminology.
- For the complete HTML template see references/学习包模板.md. Generation command:bash {SKILL_DIR}/scripts/html2png.sh 思维导图.html "00-学习路径.png" 1180 1040。
01-System Materials: Its positioning isresource library + glossary quick reference. Break the topic into knowledge modules (one sentence per module), attach an authoritative resource list to each module (title + real URL + type + one-sentence reason for recommendation); at the end, a glossary quick-reference table,one-line definition per entry(for quick lookup, not expanded into paragraphs; expansion is 03's job). Resources should be specific by name; if you cannot find it, say you cannot find it. **Do not write stage timelines, do not write long concept explanations.** For beginner level, put only the 2 to 4 most worthwhile resources in each module, prioritizing beginner-friendliness, and do not pile up expert resources.
02 Curated Video List: Each entry must include 视频标题 / 链接 / 博主 / 粉丝量(订阅量) / 该视频互动量 / 关注理由 / 对应学习阶段. Arrange in learning-path order, with real clickable links.
03-Learning-Notes.md：A Markdown document produced directly by this skill, without relying on ebook-maker. Its positioning isan ebook-style main text that explains the knowledge thoroughly: content first, using plain language + concrete examples to explain every core piece of knowledge clearly, with a preface/chapters/conclusion, at least one example per chapter, quoting authoritative original words (bilingual) at key points and attaching that URL on the spot, able to be understood independently without the original links.
- Format:Markdown（.md). The layout is not written in the MD,the ebook layout is applied when converting to PDF(see Stage 5, referring to ebook-maker's layout.md: A4 / body text 680px / primary color sepia #8B5A2B + orange accent #D35400 / serif headings).
- Do not insert illustrations / do not generate images / do not leave image placeholders.
- Markdown writing should restrain vertical bars: callout boxes/blockquotes are only used when quoting authoritative original words >(at most one place per chapter); for each chapter summary, directly use one sentence of body text without wrapping it in a box; do not wrap examples, key points, and quotes all into blocks, otherwise they will become a row of vertical lines in the PDF (this is something Irene explicitly rejected). Examples can be presented with body text + ordered lists.
- Chapters use knowledge modules as the skeleton (preface → each core knowledge block as a chapter → conclusion).
- Difficulty follows the learning package difficulty from Stage 1: beginner level first explains life scenarios and intuition, then gives terminology, using few abbreviations; intermediate level can directly use core terminology, but each chapter still needs examples; advanced level can discuss complex mechanisms, source code/papers/industry disagreements.
- Level of detail follows the Stage 1 choice(default moderate-thick):
  - Moderate-thick (default): each chapter = 2 to 3 concept subsections (###, each with one example) + one scenario demonstration + 1 to 2 common pitfalls + one sentence chapter summary.
  - Detailed tutorial version: based on moderate-thick, each chapter is split into more subsections, concepts are explained thoroughly, runnable code walkthroughs are added, common errors and troubleshooting are added, and end-of-chapter mini exercises (self-test) are added, with length about 1.5 to 2 times.
- Do not create a separate resource list, do not do stage planning(that is the job of 01 and 00).

All Chinese + bilingual original text

All narrative paragraphsentirely in Chinese。
When quoting English official/original text, use bilingual: English original sentence + Chinese translation immediately following.
Reusable commands / code / prompts retain the original language, with a Chinese explanation immediately following.

Subfolder conflict handling (automatic)

Does not exist → create directly
Exists → add for continuation -v2 / -v3; for reruns on the same day add -重跑HHMM
Strictly forbidden to overwrite old outputs

Naming rules

Topic name: Chinese or English both acceptable, spaces replaced with hyphens
Date:date +%Y%m%d taken in real time, not read from memory
Path fixed ~/Downloads/学习包/

Stage 5: Convert to PDF according to the decision made before starting (optional)

Whether to convert to PDF was already asked in Stage 1; hereexecute according to the decision made at that time：

The user said before startingconvert → convert directly after writing to disk, and tell the user the path after conversion.
What the user says before startingNo need → do not convert, wrap up.
In case Stage 1 missed asking (which should not happen), add a question here; convert only if requested.
00 is already an image and does not need conversion; only convert the ones the user wants among 01/02/03.
For conversion, use this skill's built-in script (Chinese renders correctly,hyperlinks remain clickable, and ebook layout is applied, pandoc → inline CSS HTML → headless Chrome printing):
```
bash {SKILL_DIR}/scripts/md2pdf.sh "01-系统资料.md" "02-优质视频清单.md" "03-学习笔记.md"
```
The PDF lands in the same directory with the same name .pdf。
Script key points: use --headless=new to print; Chrome will write <a> as the PDF's /Link /URI annotations (hyperlinks have been verified to be preserved); the CSS inlines an ebook style (A4 / 680px / sepia primary color / serif headings / restrained callout boxes), and Chinese falls back to system fonts so it works offline.
Do not attach other skills, conversion is a built-in capability of this skill.

Stage 6: Delivery self-check (must be completed before wrapping up)

Do a lightweight self-check before delivery; if anything fails, rework it and do not hand a half-finished product to the user:

Files complete: the directory must contain 00-学习路径.png、01-系统资料.md、02-优质视频清单.md、03-学习笔记.md. 00 is a PNG image, 01/02/03 are Markdown; 4 Markdown files are no longer required.
00 image readable: the PNG file is non-empty, and the canvas content fits on one page; include only the learning package difficulty, stage names, duration, concept keywords, and pass criteria; do not include URLs or long concept explanations.
01/02/03 do not overlap roles: 01 only does the resource library and terminology quick reference, 02 only does the video catalog, 03 only covers the knowledge body; do not start a separate resource list or make a learning path table.
Video fields complete: each video in 02 must have a title, real link, creator, follower count or subscriber count, engagement, reason to follow, corresponding stage, platform, data status, and data retrieval time. If it cannot be found, write "not obtained"; if the data is an approximate platform figure, mark it "platform approximate"; do not write an approximate figure as an exact value.
Difficulty consistent: 00/01/02/03 must all indicate and follow the learning package difficulty from Stage 1. Beginner level must not pile on expert terminology, papers, source code, and advanced paths; intermediate/advanced levels must also keep the stages clear.
PDF decision consistent: only generate PDFs for 01/02/03 if Stage 1 chose to convert to PDF; if the choice is no need, do not convert. If Stage 1 missed asking, ask again before wrapping up and then execute.
Naming and overwrite: a new directory must not overwrite an old directory; a continuation on the same topic must add -v2 / -v3 or -重跑HHMM。

Hard rules (violation means failure and mandatory rework)

Each video list entry must include creator follower count/engagement/reason to follow + real link, and data that cannot be found must be written as "not obtained"; do not make it up
The learning path must be divided into stages and each stage must have "how to verify you have learned it", and must not just pile up resources 2.5. 00 is a one-page mind map PNG image (HTML layout + html2png.sh screenshot), 01/02/03 are Markdown, all produced directly by this skill without attaching downstream skills. When converting 01/02/03 to PDF, apply ebook layout (refer to ebook-maker layout.md); callout boxes/blockquotes in Markdown should be restrained (do not lay them out as a bunch of vertical bars). The 00 mind map should look good, fit on one page, and not be crowded 2.55. The four deliverables do not overlap: In-depth conceptual explanations only in 03, one-line terminology quick reference only in 01, resource URL library only in 01, stage sequence and pacing only in 00. 00 does not expand on concepts or lay out URL lists, and 03 does not start a separate resource list or do route planning. Finding the same content duplicated in two documents = fail; put it back in its place according to the division-of-labor table. 2.6. Ask all preferences at once before starting: Before dispatching any Agent, the following must already be clearly specified: ① learning package difficulty (beginner/intermediate/advanced, default beginner) ② video platform (prefer YouTube/YT+Bilibili both/Bilibili only; must not default to deciding for the user) ③ note detail level (moderate-plus-thicker default / detailed tutorial version) ④ whether to convert to PDF. Anything the user has already stated is exempt from asking; anything not stated must be clarified in Stage 1, and only start after it has been asked. Starting without clarifying = violation.
Every key resource must have a real URL attached, do not hallucinate links, do not fabricate book titles/tutorial names
Agent minimum: Stage 2 parallel Agents < 3 → dispatch more
Videos must be attached to the corresponding stage of the learning path, they cannot be disconnected from the path and listed as a running account 5.5. Convert PDF only when the user wants it: Stage 1 must clarify whether to convert to PDF, and Stage 5 executes according to the decision at that time; if not needed, do not convert and do not make decisions on one's own.
Subfolder: Stage 4 must create a subfolder + 1 PNG image + 3 Markdown outputs; on conflict add a suffix rather than overwrite
Stage 1 receipt: even in quick mode, an inference receipt must be sent before dispatching Agents
Resource restraint: each stage gives only the minimum sufficient resources, do not dump a bunch of links to gather dust 8.5. Difficulty must be grounded: Stage 1 must list the learning package difficulty; default beginner. When generating, all materials, videos, paths, and study notes must be selected according to that difficulty, and the beginner level must be understandable to an ordinary person.
Delivery self-check: Stage 6 self-check must pass, especially checking whether 00 is really a PNG, whether 02 video fields are complete, and whether 03 has no route table or resource list

Global prohibitions

❌ Do not fabricate video links / blogger data / book or tutorial names / URLs
❌ Do not treat a single anonymous post as a "recognized learning path"
❌ Do not pile up a bunch of resources without sorting (must be sorted by stage, videos attached to stages)
❌ Do not write nonexistent materials just to "look complete"
❌ Do not repeat the content of references in SKILL.md
❌ Do not overwrite old outputs (on conflict always add a suffix)
❌ Do not take over the work of student-researcher (papers/literature reviews/citation formats/submitting Word), deep-researcher (topic selection/competitors), or irene-content-engine (writing articles)
❌ Avoid dash-style writing in narrative body text, and do not overuse dash-type symbols; URLs, code, commands, Markdown table syntax, file names, and English proper nouns remain as-is
❌ Eliminate negative-contrast paired sentence patterns: do not first negate and then turn to affirm; instead, change to a positive statement that directly states the conclusion. This is a writing flaw Irene has repeatedly pointed out, and all outputs (especially 03 study notes) must avoid it

Quick start

"帮我搜一份 Loop Engineer 的学习包，零基础想系统了解"
→ Stage 1 快捷模式 + 推断回执 → 三路并行调研 → 落 4 份产物到 ~/Downloads/学习包/20260615-Loop-Engineer/

"我想系统学 Prompt 工程"
→ Stage 1 问 4 个参数（或推断回执）→ 落 4 份产物

"做个咖啡手冲的学习包，只要中文资料，碎片时间学"
→ 跳问答，按"中文 / 碎片 20 分钟"调资源 → 落 4 份产物

"帮我学 XX 写成一篇公众号"
→ 命中 irene-content-engine 边界 → 让位