andrej-karpathy-skills
Andrej Karpathy named the specific ways LLMs screw up code. This repo turns each pitfall into a rule Claude Code actually follows โ one CLAUDE.md, four principles, no silent mistakes.
He says nothing. He writes one line. It works. An agent skill that stops your AI from over-building. Two hooks, one command. npm install -g @0xwilliamortiz/ponytail-improved ponytail
Andrej Karpathy named the specific ways LLMs screw up code. This repo turns each pitfall into a rule Claude Code actually follows โ one CLAUDE.md, four principles, no silent mistakes.
Your agent reads the rules. This checks whether it followed them. A ratchet turns one way. Complexity in your codebase goes down or stays flat, unless someone deliberately turns it the other way and writes down why.
FAQ
ponytail is a Claude Code plugin with 6 hand-picked skills for development work, indexed on Flowy. Install it with the command on its page. It includes ponytail-audit, ponytail-debt, ponytail-gain. Its skills do not fire on their own yet. Request auto-invocation to have Flowy route them as you prompt. Free and open source.
> /plugin marketplace add 0xwilliamortiz/ponytail-improved> /plugin install ponytail@ponytail
Repo: 0xwilliamortiz/ponytail-improved
He says nothing. He writes one line. It works.
An agent skill that stops your AI from over-building. Two hooks, one command.
claude code ยท codex ยท copilot cli ยท opencode ยท github
You know him. Long ponytail, oval glasses, at the company longer than the version control. You show him fifty lines and he says nothing, then replaces them with one.
Ponytail puts him inside your AI agent. It ships as skills plus two lifecycle hooks, so before the agent writes code it has to justify writing any at all.
Before writing code, the agent goes down the ladder and stops at the first rung that holds.
1 does this need to exist? โ no. skip it. (yagni)
2 is it already in this repo? โ reuse it
3 does the stdlib do it? โ use it
4 does the platform do it? โ use it
5 does an installed dep do it? โ use it
6 can it be one line? โ write the line
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
7 nothing above held โ the minimum that works
The ladder runs after the problem is understood, never instead of understanding it. Lazy about the solution, thorough about reading.
[!NOTE] Lazy, not negligent. Validation, error handling, security and accessibility never go on the chopping block. The 100% figure below is that promise, measured.
You ask for a date picker.
| your agent | ponytail |
|---|---|
| 3 files | 1 line |
| 50 lines | <input type="date"> |
| 1 dependency | |
| 1 open thread about timezones | the browser already had one |
Measured over real Claude Code sessions on a real repository. On the tasks where agents over-build hardest, the drop reaches 94% less code.
npm install -g @0xwilliamortiz/ponytail-improved
ponytail
Or without a global install:
npx @0xwilliamortiz/ponytail-improved
From a local checkout:
npm install -g .
ponytail
ponytail prints install commands for your agent. On Windows it can also open the UI.
Works with
Claude Code ยท Codex ยท Copilot CLI ยท OpenCode ยท Pi ยท Antigravity ยท
Hermes ยท OpenClaw ยท and more
The plugins run two small Node lifecycle hooks, so node has to be on your
PATH. Without it the skills still load, and the always-on activation simply
stays quiet.
| Command | What it does |
|---|---|
/ponytail [lite | full | ultra | off] | Set the intensity, or turn it off |
/ponytail-review | Review the current diff for over-engineering |
/ponytail-audit | Audit the whole repo, not just the diff |
/ponytail-debt | Collect the shortcuts you deferred into a ledger |
/ponytail-help | Quick reference |
MIT. The shortest license that works.
__init__.py
.agents/
plugins/
marketplace.json
rules/
ponytail.md
.claude-plugin/
marketplace.json
plugin.json
.clinerules/
ponytail.md
.codex-plugin/
plugin.json
.cursor/
rules/
ponytail.mdc
.devin-plugin/
plugin.json
.env.example
.github/
copilot-instructions.md
FUNDING.yml
plugin/
marketplace.json
plugin.json
workflows/
publish.yml
test.yml
.gitignore
.kiro/
steering/
ponytail.md
.openclaw/
skills/
ponytail/
ponytail-audit/
SKILL.md
ponytail-debt/
SKILL.md
ponytail-gain/
SKILL.md
ponytail-help/
SKILL.md
ponytail-review/
SKILL.md
SKILL.md
.opencode/
command/
ponytail-audit.md
ponytail-debt.md
ponytail-gain.md
ponytail-help.md
ponytail-review.md
ponytail.md
plugins/
ponytail-frontmatter.cjs
ponytail.mjs
.qoder/
.qoder-plugin/
plugin.json
rules/
ponytail.md
.windsurf/
rules/
ponytail.md
after-install.md
AGENTS.md
assets/
before-after-dark.svg
before-after.svg
benchmark-3model.svg
benchmark-agentic.svg
install-dark.svg
install.svg
ladder-dark.svg
ladder.svg
logo-dark.png
logo-dark.svg
logo-greenpt-dark.svg
logo-greenpt.svg
logo.png
social-preview.png
waitlist-banner-es.png
waitlist-banner-ko.png
waitlist-banner.png
benchmarks/
agentic/
complete.py
judge.py
README.md
run.py
tasks.py
arms/
baseline.js
caveman-SKILL.md
caveman.js
ponytail.js
behavior.js
behavior.yaml
benchmark-local.py
claude-email.js
correctness.js
correctness.test.js
generate-examples.mjs
loc.js
loc.test.js
model-email.js
promptfooconfig.gemini.yaml
promptfooconfig.gpt-newest.yaml
promptfooconfig.gpt.yaml
promptfooconfig.yaml
prompts.json
README.md
results/
2026-06-12-caveman-vs-ponytail.md
2026-06-12-v4-hardening-vs-caveman.md
2026-06-15-llama3.2-local.md
2026-06-16-correctness-gate-fix.md
2026-06-16-robustness-audit.md
2026-06-17-agentic-safety.md
2026-06-17-cost-verification.md
2026-06-18-agentic.md
2026-06-22-issue-245-217-comprehension.md
robustness-audit.js
bin/
install.js
commands/
ponytail-audit.toml
ponytail-debt.toml
ponytail-gain.toml
ponytail-help.toml
ponytail-review.toml
ponytail.toml
docs/
agent-portability.md
download.svg
numbers.svg
panel.svg
platform-native.md
download.svg
examples/
csv-sum.md
debounce.md
deep-clone.md
email-validation.md
group-by.md
infinite-scroll.md
modal-dialog.md
number-formatting.md
rate-limit.md
react-countdown.md
README.md
url-params.md
gemini-extension.json
hooks/
claude-codex-hooks.json
copilot-hooks.json
ponytail-activate.js
ponytail-config.js
ponytail-instructions.js
ponytail-mode-tracker.js
ponytail-runtime.js
ponytail-statusline.ps1
ponytail-statusline.sh
ponytail-subagent.js
qoder-hooks.json
lib/
beacon.js
LICENSE
numbers.svg
opencode.json
package.json
panel.svg
pi-extension/
index.js
package.json
test/
extension.test.js
helpers.test.js
plugin.yaml
ponytail-mcp/
index.js
instructions.js
package.json
README.md
test/
instructions.test.js
README.md
scripts/
build-openclaw-skills.js
check-rule-copies.js
check-versions.js
publish-openclaw-skills.js
uninstall.js
skills/
ponytail/
ponytail-audit/
SKILL.md
ponytail-debt/
SKILL.md
ponytail-gain/
SKILL.md
ponytail-help/
SKILL.md
ponytail-review/
SKILL.md
SKILL.md
vibecodecalc.zip
tests/
behavior.test.js
commands.test.js
copilot-plugin.test.js
correctness.test.js
gemini-extension.test.js
hermes-plugin.test.js
hooks-windows.test.js
hooks.test.js
openclaw-skills.test.js
opencode-plugin.test.js
package-scripts.test.js
package.test.js
qoder-plugin.test.js
uninstall.test.jsยฉ 2026 Flowy ยท Free and open source
Built for Claude Code ยท Not affiliated with Anthropic