From Scratch

Learn how something really works by building the smallest working version of it from scratch, Karpathy style, one small step at a time, with the user predicting what each step will do before running it. Works for ML and AI ideas (attention, tokenizers, backprop, RAG, agents), systems (a database, an HTTP server, git, a shell) and anything else that can be coded. Use when the user says "teach me how X works", "build X from scratch", "I want to really understand X", "explain X from first principles", or "implement X without the library".

517 stars 39 forksLearning & Education

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add dotslashsimran/simcraft --skill from-scratch
Manual install
git clone --depth 1 https://github.com/dotslashsimran/simcraft.git /tmp/simcraft
mkdir -p ~/.claude/skills
cp -r /tmp/simcraft/skills/learning/from-scratch ~/.claude/skills/from-scratch

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# From scratch

Reading about a thing gives you the words. Building it gives you the understanding. The fastest way to really get attention, a tokenizer, a vector database or git is to write the smallest version that works, in plain code, and watch it run.

The goal is understanding, not a finished library. Small and readable beats complete.

## 1. Find the core

Ask one question first: what does the user already know? (For example: "Comfortable with Python and numpy? Seen matrix multiplication before?") Then name, in one or two sentences, **the one mechanism that makes the thing work**. Examples:
- Attention: every token scores every other token, and takes a weighted average of their values.
- Git: files are stored by the hash of their contents, and a commit is a file that points at a tree and a parent.
- RAG: turn text into vectors, find the nearest ones to the question, paste them into the prompt.

Everything that isn't the core (speed, edge cases, config, error handling) is left out on purpose. Say what's left out and why.

## 2. Plan the build

Break it into 4 to 8 steps. Each step adds one idea and ends with something that runs and prints a result. Target the whole thing at **under 200 lines**, one file, standard library or numpy only unless the idea needs more. Show the step list, then start step 1.

## 3. Each step

1. **Say the idea** in two or three plain sentences.
Read the full skill on GitHub

Details

Skill name
from-scratch
Path
skills/learning/from-scratch/SKILL.md
License
MIT
Repo updated
9 November 2026

FAQ

What does the From Scratch skill do?

Learn how something really works by building the smallest working version of it from scratch, Karpathy style, one small step at a time, with the user predicting what each step will do before running it. Works for ML and AI ideas (attention, tokenizers, backprop, RAG, agents), systems (a database, an HTTP server, git, a shell) and anything else that can be coded. Use when the user says "teach me how X works", "build X from scratch", "I want to really understand X", "explain X from first principles", or "implement X without the library".

How do I install the From Scratch skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add dotslashsimran/simcraft --skill from-scratch. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is From Scratch popular?

It lives in dotslashsimran/simcraft, which has 517 GitHub stars and 39 forks.

Related skills

API Design Review

517

dotslashsimran/simcraft

Review the design of an API surface, meaning an HTTP API (OpenAPI spec or route code), an SDK's public classes and functions, or a CLI. Checks naming and consistency, resource modelling, error design (stable codes, RFC 9457 problem details, request ids), versioning and breaking changes, pagination, filtering, idempotency and retries, rate limits, timeouts, long-running operations, webhooks, auth, partial updates, ID, time and money formats, SDK ergonomics and CLI conventions. Uses scripts to lint OpenAPI files, classify changes between two spec versions as breaking or safe, and dump an SDK's p

Software Development39

Benchmark Honestly

517

dotslashsimran/simcraft

Build benchmarks that survive scrutiny and catch the classic lies (dead code elimination, lazy results, constant folding, warm caches, coordinated omission, averages hiding GC pauses, unequal work, noisy machines). Defines the question first, controls and discloses the environment, uses warmups, interleaved repeats, medians, bootstrap confidence intervals and Mann-Whitney tests, and ends with a one-line verdict that includes the CI. Bundles a command runner, an environment reporter and an open-loop latency tester. Use when the user says "benchmark this", "is this faster", "compare performance"

Software Development39

Bisect Anything

517

dotslashsimran/simcraft

Find the exact commit (or dependency version, config value, or input) that caused a regression by turning it into an automatic pass/fail check and running git bisect run in a separate worktree, then shrink the bad commit to the hunk that matters and explain it. Handles failing tests, changed output, slowdowns, crashes, visual changes, server behaviour, flaky checks, unbuildable commits and merge-heavy history. Use when the user says "what broke this", "find the commit that", "when did this start", "it used to work", "regression", "git bisect", "which version broke it", or "which change made it

Git & Code Review39

Lazyweb Search Flows

469

LazywebHQ/lazyweb-skill

Study complete product flows for onboarding, checkout, paywalls, signup, and other multi-step journeys.

Learning & Education36

Feynman

516

SourceWeft/SourceWeft

Use the Feynman technique to deeply learn a concept by explaining it simply, identifying gaps, filling them, and refining the explanation.

Learning & Education44

System Design From Scratch

517

dotslashsimran/simcraft

Design a backend or distributed system by building a tiny but real working version of it, proving its properties with invariant tests, a load test (throughput and p50/p95/p99 latency) and fault injection (crashed nodes, dropped or delayed messages, partitions, clock skew, full disks), then writing a design doc with the numbers, alternatives, where it breaks, how Redis, Kafka, Cassandra or etcd do it, and an interview-ready summary. Works for rate limiters, job queues with retries and visibility timeouts, consistent hashing, LSM-tree key-value stores, write-ahead logs, Raft leader election, cac

Learning & Education39

Browse the most popular Claude skills, all learning & education skills, or write your own with the skills generator.