Works with the AI Agents you already use
Knowledge Adapter
Turns one sentence — "I want to learn diffusion models" — into a curated set of papers, tutorials and repos, then renders a self-contained multi-chapter HTML course with MathJax derivations, quizzes and dark-mode toggle.
Free
knowledge-adapter
Example session with this skill installed
I want to properly learn video generation models — I've seen Sora demos but I don't
actually understand how any of this works under the hood.
Can you collect the foundational papers on video generation, plus a hands-on tutorial
and the main open-source implementation, then build me a tutorial I can work through
end to end? I'm comfortable with calculus and linear algebra but I haven't done
probability in a while, so tell me upfront what I should brush up on rather than
teaching it to me.
I want to actually be tested on whether I understood it, not just read through it.
- Read your context and instructions
- Compiled the knowledge-adapter
Collected 7 materials for "video generation models"
| # | Type | Material | Why this one |
|---|---|---|---|
| 1 | Paper | Video Diffusion Models (Ho et al., 2022) | Foundational — first general video diffusion framework |
| 2 | Paper | Make-A-Video (Meta, 2022) | Trains text-to-video without paired text-video data |
| 3 | Paper | DiT: Scalable Diffusion with Transformers (2023) | Replaces U-Net with a transformer backbone; the architectural basis for Sora |
| 4 | Paper | Align your Latents / Video LDM (2023) | Latent-space video diffusion; precursor to Stable Video Diffusion |
| 5 | Paper | AnimateDiff (2023) | Image-to-video via plug-in motion module; most practical of the five |
| 6 | Tutorial | Text to Video Generation with HuggingFace Diffusers | Runnable AnimateDiff walkthrough |
| 7 | Repo | guoyww/AnimateDiff (12.2k stars) | Official implementation — README + core source, shallow clone |
Downloaded as arXiv e-print source packages: 5 × .tex + .bib + figures/,
not PDFs. Output at target/papers/, target/tutorials/, target/repos/.
Manifest: harvest_manifest.json.
Built a 10-chapter course — tutorial/
Prerequisite self-check (listed on the index page, not taught):
| Subject | You should be comfortable with |
|---|---|
| Calculus | Derivatives, gradient ∇, chain rule, partial derivatives |
| Probability | Random variables, Gaussian distribution, expectation & variance, conditional probability, sampling |
| Linear algebra | Vectors, matrices, matrix multiplication, norms |
| ML basics | Loss functions, gradient descent, backpropagation |
Chapter outline
| # | Chapter | Level | Time |
|---|---|---|---|
| 1 | Video generation primer: task, pipeline, symbol table | 🟢 | 30 min |
| 2 | Diffusion fundamentals: DDPM and denoising score matching | 🟢 | 32 min |
| 3 | Latent diffusion and the U-Net backbone | 🟢 | 32 min |
| 4 | Transformers, ViT and LoRA | 🟢 | 32 min |
| 5 | VDM: a space-time factorized 3D U-Net | 🟡 | 34 min |
| 6 | Make-A-Video: pseudo-3D without video data | 🟡 | 32 min |
| 7 | DiT: replacing the U-Net with a transformer | 🟡 | 32 min |
| 8 | Video LDM: temporal layers over a frozen image model | 🟡 | 34 min |
| 9 | AnimateDiff: motion module and LoRA | 🔴 | 32 min |
| 10 | Comparing the five approaches | 🟣 | 20 min |
Total ~310 minutes.
What each chapter contains
第5章: 视频扩散模型 VDM——时空分解的 3D U-Net ← nav: prev / index / next / 🌙
⏱ ~34 分钟 🟡 进阶 前置: 第2、3章
📝 前置小测 — 3 questions, each with a toggle answer
Q1: 为什么把 2D 卷积改成 1×3×3 就能处理视频?
Q2: 时序注意力与空间注意力的 batch 轴有何不同?
📖 教学内容
5.1 从图像扩散到视频扩散
The 2D U-Net from Chapter 3 processes one frame. To handle a video we
keep the spatial path unchanged and add a second, orthogonal path over time.
$$(f * g)[i,j] = \sum_m \sum_n f[i+m, j+n] \cdot g[m,n]$$
💡 Key point — factorized space-time attention: spatial attention treats
the frame axis as batch; temporal attention treats the spatial axes as
batch. Cost stays linear in frames instead of quadratic.
5.2 联合训练图像与视频
⚠ Masking the temporal attention turns the model back into an image model,
which is what makes joint image-video training possible at all.
🔍 Reconstruction-guided sampling (click to expand)
Step 1 — the replacement method substitutes exact forward-process
samples for the conditioning block…
Step 2 — but the score term ∇ log q(x^a | z_t) is missing, so the two
blocks drift apart…
Step 3 — approximating that term with a Gaussian gives
$$\tilde{x}^b_\theta(z_t) = \hat{x}^b_\theta(z_t)
- \frac{w_r \alpha_t}{2}\nabla_{z^b_t}\|x^a - \hat{x}^a_\theta(z_t)\|^2$$
Concrete check: FVD 295 on UCF101 vs 3431 for TGAN-v2 — a 11.6× gap.
✅ 后置自测 — 4 questions, each with answer + explanation of why
Files
tutorial/
├── index.html learning path, prerequisite self-check, chapter cards
├── chapter_01.html … chapter_10.html
├── json/ content JSON per chapter — enables free re-rendering
└── (396 KB total)
Open tutorial/index.html in a browser — no server, no build step.
Next steps
- Open the index page first. Check the prerequisite table against what you
actually remember. If probability is shaky, brush up before Chapter 2 — the
diffusion derivation assumes you're comfortable with Gaussians and expectation. - Work chapters 1→10 in order. The dependency chain is real: Ch.5 assumes Ch.2
and Ch.3; Ch.10 assumes you've read 5–9. - Don't skip the quizzes. The pre-quiz tells you whether a chapter is worth
your time; the post-quiz explanations are where the "why" lives. - Want different coverage? Edit
generation_input.jsonand re-run
gen_chapters.py. Only want to restyle?rerender.pyrebuilds all chapters
from saved JSON with no API calls.
---
Connects securely to your tools. The creator never sees your data.
About this skill
Knowledge Adapter
Most of the work in learning a new field isn't the learning — it's the material wrangling before it. Which papers actually matter? Where does this one paper sit in the field? What prerequisite am I missing? And after reading, how do I check whether it stuck?
Knowledge Adapter automates both halves of that problem with two coordinated skills.
What it does
Stage 1 — topic-harvest (collect).
Give it a topic in one sentence. It searches arXiv, Semantic Scholar and OpenAlex
via their public APIs, ranks results by citation count, and downloads the arXiv
e-print source packages (.tex + .bib + figures) rather than PDFs — much cleaner
input for downstream processing. It also finds hands-on tutorials and shallow-clones
the canonical open-source implementation (README + core source, not the whole repo).
Collection is deliberately bounded: 4–6 papers, 2–4 tutorials, 1–2 repos. You get a
ranked manifest to approve before anything expensive happens.
Stage 2 — knowledge-adapter (teach).
It reads the materials in full, extracts the concept graph, and — this is the part
most tools skip — identifies prerequisites the source material uses but never
explains. Those become dedicated opening chapters. The syllabus is topologically
sorted and each chapter is capped at 20–35 minutes of reading.
Every chapter follows a fixed three-part structure: pre-quiz (should you even read this?) → teaching content → post-quiz (with explanations of why, not just what).
What you get
Self-contained single-file HTML. CSS and JS inlined, MathJax the only CDN call. Double-click to read. Light/dark toggle in the nav bar, choice persisted in localStorage with no flash on reload. Prints cleanly.
A real generated example (10 chapters on video generation models) ships in the repo
at examples/video-generation-models/tutorial/index.html.
Design decisions worth knowing
Content and presentation are decoupled. The LLM produces only structured
content JSON — block types are pinned by schema. A deterministic Python renderer
(render_chapter.py) is the only thing that writes HTML. This exists because an
earlier version let each chapter be written independently by an agent, and the
chapters drifted apart: different CSS systems, different MathJax configs, different
quiz markup. Now consistency is structural, not a matter of good behavior. Change a
style once, every chapter updates.
Your math background is your business. The course does not re-teach probability,
calculus, linear algebra or ML fundamentals — that's a textbook's job. Instead the
index page lists a "prerequisite self-check" (e.g. probability — Gaussian
distribution, expectation, conditional probability) so you know what to brush up
on. The course itself covers only domain knowledge: task definition, the abstract
pipeline, the research landscape, and a symbol table so you know what x_t means
when you meet it.
API keys never touch the command line. CLI arguments leak into shell history and
process lists. Keys come from environment variables or a chmod 600 config file —
--api-key deliberately does not exist. Any OpenAI-compatible endpoint works; it is
not tied to one vendor.
Re-rendering is free. Chapter content JSON is saved to tutorial/json/. After
you tweak the renderer, rerender.py rebuilds every chapter with zero LLM calls.
Requirements
Claude Code · Python 3.8+ · any OpenAI-compatible API (DeepSeek, OpenAI, OpenRouter,
or local Ollama). pip install -r requirements.txt — only requests is required,
json5 optional.
Honest limitations
- Phases 1–3 (material analysis, knowledge graph, syllabus design) run inside the Claude Code conversation, not as standalone scripts. Phases 4–5 are standalone Python. The full pipeline needs Claude Code.
- Generation quality tracks the model you configure. Reasoning models produce noticeably better derivations but cost more.
- Material relevance and download success rate have not been systematically benchmarked.
MIT licensed.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse All Claude Skills.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 17 days ago
- Free with an Agensi account
- Accounts
GitHub
OpenAI
Arxiv
Deepseek
- Software
Git
Permissions
Allowed Hosts