Portable skills for building better and faster with coding agents like Codex, Claude Code, and other SKILL.md-compatible tools.
These skills are inspired by the workflow Andrej Karpathy uses to build quality software faster and more reliably with AI agents: write a precise spec, define a verifier, and improve the environment the agent works inside.
This collection is intentionally small. V1 focuses on one reusable engineering loop plus a research-only ask mode:
- Turn vague work into a tight spec.
- Define verification before implementation.
- Build inside a reusable agent environment.
- Run the work in small checkpoints instead of one giant handoff.
After this repository is published on GitHub:
npx skills add ipenywis/skillsThis installs the collection from
ipenywis/skills.
| Skill | Purpose |
|---|---|
c1-agent-build-loop |
Orchestrate a full spec-to-verify coding-agent session. |
c1-agent-spec |
Convert vague requests into decision-ready implementation specs. |
c1-agent-verify |
Create and run verification plans for agent-built work. |
c1-agent-environment |
Set up reusable project context, knowledge layout, and guardrails. |
c1-ask |
Investigate, debug, and answer directly without mutating until explicit approval. |
skills/
engineering/
c1-agent-build-loop/
SKILL.md
c1-agent-spec/
SKILL.md
c1-agent-verify/
SKILL.md
c1-agent-environment/
SKILL.md
c1-ask/
SKILL.md
This collection highlights the Karpathy-inspired workflow for high-quality agent-assisted software development: make the goal explicit, turn it into a tight spec, verify the result with real signals, and compound the environment over time. These are not official skills from Karpathy, any model provider, or any public speaker.
Keep each skill portable and concise:
- Use only
nameanddescriptioninSKILL.mdfrontmatter. - Keep workflow instructions in the skill body.
- Add resource folders only when a skill genuinely needs them.
- Require explicit written approval before edits, installs, migrations, commits, deploys, deletes, or other mutating work.