Karpathy Loop

Use when building self-improving autonomous systems, optimization loops, or applying experiment-driven iteration to any domain. Triggers on "auto-improve", "self-optimizing", "experiment loop", "karpathy loop", autonomous optimization, or when a workflow could benefit from measure-analyze-rewrite cycles.

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add artguru-tech-team/claude-skills-public --skill karpathy-loop
Manual install
git clone --depth 1 https://github.com/artguru-tech-team/claude-skills-public.git /tmp/claude-skills-public
mkdir -p ~/.claude/skills
cp -r /tmp/claude-skills-public/karpathy-loop ~/.claude/skills/karpathy-loop

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

From the SKILL.md

# Karpathy Loop - Autonomous Experiment Framework

Turn any measurable workflow into a self-improving loop. Based on Karpathy's autoresearch (700 ML experiments in 2 days) and proven adaptations to content, outreach, and ops.

## Core Loop

```
Seed Knowledge --> Generate Hypothesis --> Run Experiment --> Measure --> Analyze --> Rewrite Strategy --> Repeat
```

## The 3-Register Program

Every loop needs a `program.md` (or equivalent) with exactly three registers:

| Register | Purpose | Example |
|----------|---------|---------|
| **Instructions** | What to search for / try next | "Test subject lines with numbers vs questions" |
| **Constraints** | What must NOT change (invariants) | "Never exceed 3 emails per sequence, keep tone professional" |
| **Stopping criteria** | When to stop and report | "After 50 experiments OR when metric plateaus for 3 rounds" |

## 6 Patterns to Apply

### 1. Seed Before You Loop
Never cold-start. Seed with 15-25 real examples of "what good looks like."
- Bad seed = wasted loops. The seed defines the search space.

### 2. Metric-Driven Self-Rewrite
The agent scrapes real-world metrics, then rewrites its own strategy file.

| Domain | Metric | Strategy file |
|--------|--------|---------------|
| ML training | loss, speed | hyperparams.yaml |
| Content | impressions, engagement | content_strategy.md |
| Cold email | open rate, reply rate | sequence_strategy.md |
| Support routing | misroute rate | routing_rules.json |
Read the full skill on GitHub

Details

Skill name
karpathy-loop
Path
karpathy-loop/SKILL.md
Repo updated
11 November 2026

FAQ

What does the Karpathy Loop skill do?

Use when building self-improving autonomous systems, optimization loops, or applying experiment-driven iteration to any domain. Triggers on "auto-improve", "self-optimizing", "experiment loop", "karpathy loop", autonomous optimization, or when a workflow could benefit from measure-analyze-rewrite cycles.

How do I install the Karpathy Loop skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add artguru-tech-team/claude-skills-public --skill karpathy-loop. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Karpathy Loop popular?

It lives in artguru-tech-team/claude-skills-public, which has 10 GitHub stars and 0 forks.

Related skills

Amplitude Measurement Plan

10

artguru-tech-team/claude-skills-public

Amplitude measurement plan for a B2B two-sided marketplace.

Other

Appels Offres

10

artguru-tech-team/claude-skills-public

Use when responding to French public procurement tenders (appels d'offres, marchés publics, MAPA, DCE). Triggers on AO analysis, mémoire technique drafting, DCE review, CRF pricing, security questionnaire, candidature preparation. Also use when evaluating Go/No-Go on a tender.

Software Development

Brand Builder

10

artguru-tech-team/claude-skills-public

Use when building a complete brand identity for a new or existing brand — brand strategy, voice document, moodboard, logo concepts, color palette, and full AI-ready brand kit. Triggers on "build my brand", "create brand identity", "brand guidelines", "brand kit", or when a project needs visual + copy foundations from scratch.

Frontend & Design

Exec Prep Note Generic

5

Gainsight-Central/gainsight-ai-cookbook

Produces a structured Executive Prep Note for a VP or C-suite joining a client call. Synthesizes Gainsight CS + Staircase AI data. Six sections: Executive Summary, Why We're Having This Call (1–2 bullets), How They Make Money (1–2 bullets), Key Strategic Objectives (O2), Key Contacts, Risk Overview (combined), Client Maturity Index. Posts to Gainsight Timeline automatically. ALWAYS trigger for: "exec prep note for [Account]", "prepare my exec for [Account]", "exec brief for [Account]", "executive prep for [Account]", "get [exec] ready for [Account]", "executive briefing for [Account]", "prep n

Productivity & Notes

Stealth Browser

5

zippoxer/claude-skills

Stealth browser automation with anti-bot bypass and persistent page state. Use when users need to navigate Cloudflare-protected sites, fill forms, take screenshots, extract web data, or automate browser workflows on sites that block regular automation. Trigger phrases include "go to [url]", "click on", "fill out the form", "take a screenshot", "scrape", "automate", "bypass cloudflare", or any browser interaction request on protected sites.

Productivity & Notes

Risk Update Skill

5

Gainsight-Central/gainsight-ai-cookbook

Generates and posts Risk Timeline Updates in a Customer Success Platform (CSP) for accounts where the operating CSM owns an active Risk task. Combines conversation intelligence (CCI) data (sentiment, churn signals, risk scores, suggested playbooks) with the most recent call transcripts to write a structured entry, including executive summary, key risk signals, call context, and Recommended Next Steps. Posts the update onto the Risk task's Timeline. **Always trigger for:** "risk update", "update my risk tasks", "post risk updates", "update at-risk accounts", "log risk notes", "write risk timeli

Productivity & Notes

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