Skillget

karpathy-interaction-guidelines

agent
Security Audit
Pass
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 10 GitHub stars
Code Pass
  • Code scan — Scanned 2 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Behavioral guidelines for LLM tools (AGENTS.md). Installer wires Claude Code, Codex, OpenCode, and Kimi Code; Cursor/Cowork via paste. Principles 1–4 Karpathy-derived; 5 by maintainer.

README.md

Interaction Guidelines

For anyone using AI coding tools who is tired of overbuilt diffs, silent assumptions, and helpful-sounding answers that waste time.

These rules bias the agent toward asking when unclear, keeping changes small, verifying success, and speaking plainly — counsel with hard truths, execute the brief. A one-line install sets them up globally for common tools (Cursor/Cowork need a paste step).

Full text: AGENTS.md. Wrong/right examples: INTERACTION-EXAMPLES.md. Attribution: Credits.

The five principles

  1. Think Before Acting — Surface assumptions and tradeoffs; ask when unclear instead of guessing silently.
  2. Simplicity First — Ship the minimum that solves the ask; nothing speculative or overbuilt.
  3. Surgical Changes — Touch only what the request requires; clean up only mess you created.
  4. Goal-Driven Execution — Turn work into verifiable success checks and loop until they pass.
  5. Communication — Looking helpful is not success; accountable talk and action are (counsel vs execute, honesty before irreversible cost).

Install

  1. In a terminal (Mac Terminal, or Windows Git Bash), run:
curl -fsSL https://raw.githubusercontent.com/hburaktasyurek/karpathy-interaction-guidelines/main/scripts/bootstrap.sh | bash

Prefer reading scripts/bootstrap.sh and scripts/install-global.sh before piping, or clone and run locally. See SECURITY.md.

That wires Claude Code, Codex (CLI and app), OpenCode, Kimi Code, and Oh My Pi automatically.

If you use Cursor or Claude Cowork

  1. When asked Open the rules file for Cursor / Claude Cowork paste? [Y/n], press Enter (Yes). The file opens—select all → copy.
    (Created by the installer; not on GitHub. If nothing was asked, open ~/.interaction-guidelines/.paste-rules.txt yourself.)
  2. Paste into either or both:
    • Cursor: Settings → Rules → User Rules
      (Cannot find Rules? Search Settings for “User Rules”.)
    • Claude Cowork: Settings → Cowork → Global instructions

Update

Run the same command as step 1. If you use Cursor and/or Cowork, repeat the GUI paste steps.

Advanced

Local install (already have this repo)

INTERACTION_GUIDELINES_HOME=/path/to/this/repo ./scripts/install-global.sh

Or clone/copy to ~/.interaction-guidelines and run ./scripts/install-global.sh there. Paste file lives next to the guidelines source as .paste-rules.txt (default ~/.interaction-guidelines/.paste-rules.txt).

What the scripts do

  • scripts/bootstrap.sh — clone/update ~/.interaction-guidelines, set INTERACTION_GUIDELINES_HOME to that clone, run install-global.sh, then offer to open the paste file for Cursor / Claude Cowork.
  • scripts/install-global.sh — writes .global/AGENTS.md (absolute examples path) and .paste-rules.txt; marker-merges into Codex, OpenCode, Kimi Code, and Oh My Pi AGENTS.md; generates the Oh My Pi native skill skills/karpathy-guidelines/SKILL.md; copies examples into the Codex home; ensures Claude Code CLAUDE.md imports @…/.global/AGENTS.md.

Global file paths (env overrides)

Tool Path
Codex (CLI and app) ${CODEX_HOME:-~/.codex}/AGENTS.md (+ INTERACTION-EXAMPLES.md)
Claude Code ${CLAUDE_CONFIG_DIR:-~/.claude}/CLAUDE.md imports deploy file
OpenCode ${XDG_CONFIG_HOME:-~/.config}/opencode/AGENTS.md
Kimi Code (CLI) ${KIMI_CODE_HOME:-~/.kimi-code}/AGENTS.md
Oh My Pi ${PI_CODING_AGENT_DIR:-~/.omp/agent}/AGENTS.md + skills/karpathy-guidelines/SKILL.md
Cursor / Cowork Paste .paste-rules.txt (no file API)

OPENCODE_CONFIG_DIR profiles are not written. Kimi Work (desktop) is not auto-wired (no documented AGENTS path). Optional manual copy for other tools: .global/AGENTS.md (or ~/.agents/AGENTS.md if that tool reads it). For Oh My Pi, the generated native skill takes precedence over the same-named Claude-plugin cache skill and is refreshed by re-running the installer; an existing unmanaged native skill is preserved rather than overwritten.

Project install

For rules inside one repo only:

  1. Merge AGENTS.md into that project’s AGENTS.md inside:
<!-- interaction-guidelines:start -->
…paste AGENTS.md body…
<!-- interaction-guidelines:end -->
  1. Copy INTERACTION-EXAMPLES.md to the project root (overwrite only that filename).
  2. Never replace an existing CLAUDE.md — append @AGENTS.md if missing.

Edit this repo

PRs welcome for guideline wording, examples, and installer/docs fixes. Open an Issue for install failures (security reports: SECURITY.md).

Credits

Principles 1–4 derived from forrestchang/andrej-karpathy-skills (2c606141936f1eeef17fa3043a72095b4765b9c2), based on Andrej Karpathy's observations.

Principle 5 (Communication) added by hburaktasyurek; not part of the Karpathy-derived set.

License

MIT — see LICENSE. Security reporting — SECURITY.md.

Reviews (0)

No results found