A GitAgent definition providing an AI pedagogue layered on pageman/sutskever-30-implementations
This repo does not contain notebooks. It defines an AI agent that reads, explains, traces, and reviews the 30 pure-NumPy deep learning implementations in the companion repository. Each skill is:
- Grounded in the original paper via arXiv fetching — no hallucinated citations
- Anchored to the actual notebook cells via GitHub API reads
- Structured with a consistent 5-part explanation format
- Governed by explicit rules (no spoilers, NumPy-only, cite sources)
- Delegatable to specialist sub-agents (math derivations, code review)
# Prerequisites: Node.js >= 18, a GitHub token
npm install -g @shreyaskapale/gitagent
git clone https://github.com/pageman/sutskever-agent
cd sutskever-agent
gitagent validate
# Export and run with Claude Code
gitagent export --format claude-code --output exports/claude-code
claude --system-prompt exports/claude-code| # | Skill | Description | Tools |
|---|---|---|---|
| 1 | explain-paper |
5-part explanation: Problem → Key Insight → Math → NumPy → Context | arxiv-fetcher, github-reader |
| 2 | explain-notebook |
Cell-by-cell walkthrough with shape annotations and gradient moment detection | github-reader, notebook-parser |
| 3 | compare-papers |
Side-by-side table + prose synthesis of thematic connections | arxiv-fetcher, github-reader |
| 4 | review-contribution |
PR review: correctness, NumPy compliance, pedagogical structure | github-reader, notebook-parser |
| 5 | generate-exercises |
3-tier exercises: Comprehension / Extension / Research | github-reader, notebook-parser |
| 6 | summarize-repo |
Navigable overview of all 30 papers filtered by track or cluster | github-reader |
| 7 | trace-gradient |
Chain rule walkthrough from loss to input with stability annotations | notebook-parser, github-reader |
| Workflow | Trigger | Description |
|---|---|---|
onboard-new-user |
First session | Assess background, assign learning track, recommend entry point |
pr-review-pipeline |
PR opened on source repo | Full technical + pedagogical review with human gate before posting |
study-session |
Any session | Guided paper study: explain → walk notebook → trace gradient → exercises → memory |
| Agent | Temperature | Role |
|---|---|---|
math-explainer |
0.2 | Step-by-step derivations — never skips steps, never hand-waves |
code-reviewer |
0.1 | CRITICAL / WARNING / SUGGESTION findings on NumPy implementations |
sutskever-agent/
├── agent.yaml # Root manifest — model, skills, tools, compliance
├── SOUL.md # Agent identity, values, communication style
├── skills/ # 7 skill definitions (explain-paper/, trace-gradient/, ...)
├── tools/ # 3 tool definitions (github-reader, notebook-parser, arxiv-fetcher)
├── workflows/ # 3 multi-step workflows
├── knowledge/ # Static reference: paper index, notation guide, NumPy patterns
├── rules/ # 4 behavioral guardrails (no-spoilers, numpy-only, ...)
├── hooks/ # Session lifecycle hooks + shell scripts
├── agents/ # Sub-agents: math-explainer/, code-reviewer/
├── compliance/ # Risk assessment, model card
└── .github/workflows/ # CI: gitagent validate on every push
The agent onboards new users into one of four tracks drawn from the source repo:
Beginner — Papers 7 → 10 → 11 → 22 → 26 → 2 → 3 (start: AlexNet, ~21 hours)
Intermediate — Papers 2 → 3 → 4 → 6 → 13 → 14 → 16 → 18 → 20 → 22 → 26 → 28 (start: Char-RNN, ~57 hours)
Advanced — Papers 5 → 8 → 9 → 12 → 17 → 19 → 21 → 23 → 24 → 25 → 27 → 29 → 30 (start: MDL, ~69 hours)
Theory — Papers 5 → 23 → 24 → 25 → 1 → 8 → 19 (start: MDL, ~35 hours)
Source repo is never modified. This agent is a pure overlay — it reads pageman/sutskever-30-implementations via the GitHub API. No notebooks are changed.
NumPy-only rule is enforced at every layer. The rules/numpy-only.md rule is applied in skill instructions, exercise generation, and PR reviews. The code-reviewer sub-agent flags any torch/jax/tensorflow import as CRITICAL.
No spoilers by default. rules/no-spoilers.md lists the core "trick" of each paper and prevents the agent from revealing it before the learner has engaged with the motivating problem.
PR review requires human approval. The pr-review-pipeline workflow includes a human gate: if any CRITICAL finding exists, the compiled review is shown to the user for approval before being posted as a GitHub comment.
Exports are generated, not committed. Run gitagent export --format <fmt> --output exports/<fmt> to generate adapter-specific versions. Supported: claude-code, openai, crewai, github.
gitagent export --format claude-code --output exports/claude-code
gitagent export --format openai --output exports/openai
gitagent export --format crewai --output exports/crewai
gitagent export --format github --output exports/github-actionsThis repo follows the GitAgent v0.1.0 spec. All PRs run gitagent validate via .github/workflows/validate.yml — the PR must pass validation before merge.
To add a skill: create skills/<hyphen-name>/SKILL.md with the required front-matter, register the name in agent.yaml under skills:, and run gitagent validate locally before opening a PR.
Commit message convention follows the companion repo:
feat:for new skills, tools, or workflowsfix:for correctionsdocs:for README and knowledge base updates
@misc{sutskever-agent-2026,
author = {Paul "The Pageman" Pajo, pageman@gmail.com},
title = {sutskever-agent: A GitAgent pedagogue for the Sutskever 30-paper reading list},
year = {2026},
url = {https://github.com/pageman/sutskever-agent},
note = {GitAgent v0.1.0 definition layered on pageman/sutskever-30-implementations}
}Educational use. See individual papers for original research citations.
- Ilya Sutskever for the original reading list
- pageman/sutskever-30-implementations — the companion repository this agent is built on
- GitAgent / @shreyaskapale for the open agent standard