What the tech

Search

/ opens search from anywhere. Esc closes it.

Tools

Get the newsletter

karpathy-guidelines: a Claude Code skill by Multica

For developers who want Claude to ask before guessing, keep changes small and prove work with tests. Skip it for quick throwaway scripts where the caution slows you down.

Score
63of 100
Rank
#218of 239
Installs
44Kon skills.sh
Security
Clear 3 scanners
Updated
2026
Context
~59tokens always

Install

npx skills add https://github.com/multica-ai/andrej-karpathy-skills --skill karpathy-guidelines

Run it in your terminal. View the source

Using the Claude app rather than Claude Code? You can add a skill there too. How to add it in the app

What does karpathy-guidelines do?

karpathy-guidelines is a skill made of four short rules for how Claude should write code, based on Andrej Karpathy's comments about common LLM coding mistakes. First, state assumptions, raise other readings of the request and ask when something is unclear. Second, write the least code that solves the problem, with no unrequested features or abstractions. Third, touch only lines tied to the request, match existing style and mention, rather than delete, dead code you notice. Fourth, turn tasks into checks, such as writing a failing test for a bug, and give a short plan with a check for each step. It runs no commands.

When should you use karpathy-guidelines?

Use it when

  • Claude keeps rewriting code you didn't ask it to touch
  • You want a test to define done before a fix
  • Requests are vague and you want questions first

Skip it when

  • You want rapid prototypes without check-ins
  • Your CLAUDE.md already has the same rules

Why it scores 63

Strongest on quality (100 of 100), weakest on upkeep (40).

  • Adoption35 points 60
  • Upkeep20 points 40
  • Trust20 points 65
  • Quality15 points 100
  • Reader answers10 points Not enough answers yet, so this does not count

How the formula works

Is karpathy-guidelines safe to install?

Every scanner that checked it passed it. Scanners often miss what other scanners catch, and a skill can change after a scan. Read its source before you install it.

ScannerResultDate
Gen Agent Trust HubPassedSep 15, 2026
SocketPassedSep 15, 2026
SnykLow riskSep 15, 2026

Details

Installs
44,045
Repository stars
231,961
Last repository change
Tokens always loaded
about 59
Tokens when used
about 556
License
MIT
Data confidence
High

Alternatives to karpathy-guidelines

All skills for this job