convez.ai
SkillProdutividadeIntermediário
quzhi-ai

Deep Learning - Tutor Socrático

quzhi-ai/deep-learning

Tutor que força compreensão profunda através do método socrático

Ver no GitHub
InicialPluginScore · de 0 a 100

Uma skill que funciona como um tutor que se recusa a dar respostas diretas até você provar que realmente compreendeu o conceito. Ensina através de reafirmações, questionamento profundo de "por quê" e verificação prática, colapsando a lacuna entre reconhecimento superficial e entendimento real.

Instalação

$npx convez-ai quzhi-ai/deep-learning

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
8
Forks
2
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Aprender conceitos técnicos e teóricos com profundidade real
  • Treinar habilidades de explicação e raciocínio crítico
  • Estudar codebases, artigos técnicos e sistemas complexos

Veredito

Uma abordagem pedagogicamente sólida que contrasta com assistentes IA convencionais. Força entendimento autêntico em vez de reconhecimento fluente, mas exige paciência e comprometimento do usuário em um processo iterativo demorado.

README

A tradução para o português deste README ainda está na fila.

🌐 English · 中文

Ask to learn anything — a concept, a system, a codebase, an article's underlying logic — and this skill won't dump the answer on you. It teaches step by step, makes you restate each piece in your own words, drills "why" down to the root cause, and tracks your progress in a live Markdown checklist.

The whole point: collapse the gap between "sounds right, I think I got it" and "I can actually explain it and name the edge cases." Most AI Q&A leaves you fluent but forgetful. This one makes you prove it.

npx skills add quzhi-ai/deep-learning

Why · How it works · Install · Example session


Why deep-learning

Ask a normal AI assistant to explain something and you get a clean, fluent answer. You nod. You move on. Three days later you can't reconstruct it — because you never actually understood it, you just recognized it.

Recognition is not understanding. The dangerous middle state is "I sort of get it" — confident enough to stop, too shallow to use or teach.

deep-learning is built to kill that middle state. Its two sharpest moves:

  1. Restate-first — before it adds anything, it makes you explain in your own words. Your restatement reveals exactly where the holes are.
  2. Why-laddering — it doesn't stop at what and how; it keeps asking why until you hit the root cause. Understanding the essence beats memorizing the conclusion.

You don't "finish" until you can restate the whole thing in your own words and produce a counter-example or edge case. No graduation by nodding.


How It Works

A teaching loop, not a lecture:

  1. Calibrate — it first probes what you already think and where you've seen the topic, so it pitches at the right depth instead of over- or under-explaining
  2. Teach one step — one idea at a time; it never spoils what comes next
  3. You restate — you explain it back in your own words
  4. Patch the gaps — it corrects misconceptions and fills holes based on your restatement, not a generic script
  5. Ladder down "why" — it keeps digging until you reach the root cause
  6. Verify — open / case / multiple-choice questions, answers never revealed up front
  7. Graduate — only when the checklist is fully checked and you can restate + give an edge case
You: "I want to truly understand X"
  → Calibrate depth → Teach one step → You restate in your words
  → Patch gaps → Ask "why" deeper → Quiz (no spoilers)
  → ✓ checklist complete + you can give an edge case → Graduate

The learning checklist (maintained throughout)

The tutor keeps a live Markdown checklist so you always see your progress:

BlockTracks
1. The problem itselfWhat it is · why it arises · what variant cases / branches exist
2. The solutionWhat it is · why designed this way · key design decisions · edge cases & special cases
3. The bigger pictureWhy it matters · how it relates to the larger system · what the downstream impact is

Explain-like-I'm levels

Ask for any depth and it adapts:

LevelStyle
ELI5Like explaining to a 5-year-old
Middle schoolAnalogies and intuition
UniversityPrinciples and derivation
New hireReal work scenarios and a path to get hands-on

Example session

A real run, learning the concept "Harness" (the scaffolding that turns a raw AI model into a system that actually gets work done):

You: I want to truly understand what "Harness" means.

Tutor: Before I explain — tell me first, so I pitch this right: where have you seen the word? What's your rough guess at its meaning? And do you know what it literally refers to as a physical object?

You: Original meaning is "horse harness." I think it means a set of constraints that stop the AI from running wild...

Tutor: The horse-harness etymology — correct. But you've grabbed the secondary half. A harness doesn't exist to restrain the horse — it exists to make the horse useful: to convert its raw power into a cart that actually moves toward a goal. Restraint is one strap; transmission and direction are the point. Now — quick case question: which of these are part of a "harness," and which one isn't...

The tutor keeps a checklist, makes you reason, corrects the exact misconception in your restatement, and only closes the session once you can define the concept in your own words and apply it to your own work.

See examples/ for full transcripts.


Install

npx skills add quzhi-ai/deep-learning

Or manually: clone this repo and copy SKILL.md into your agent's skills directory:

cp -r deep-learning ~/.claude/skills/

Then just say:

"I want to truly understand [topic]." · "Teach me [X] for real, not just the answer." · "我想搞懂 [X]" · "学透 [X]"

The tutor takes it from there.


File Structure

deep-learning/
├── SKILL.md        # The skill definition (read by Claude / any agent)
├── README.md       # English (this file)
├── README.zh.md    # 中文
├── examples/       # Full learning-session transcripts
└── LICENSE         # MIT

It's a single-file skill — no scripts, no dependencies. Just a teaching protocol.


When NOT to use it

deep-learning is for understanding — not for doing, deciding, or fixing:

  • Pure fact lookups ("what day is it", "what's this API call") — just answer directly
  • Fast execution (write an email, build a deck, ship a report) — use a task tool
  • Decisions (should I do X?), design (build a system from scratch), diagnosis (find the root cause of a failure) — use the appropriate specialized tool

It does one thing: turn "I think I get it" into "I can prove I get it."


Star History

License

MIT — see LICENSE


Ver o original no GitHub

Bom para

Relacionados