Agent Friendly Code

Agent Friendly Skill: score your repo, pick a model

An agent skill that scores the repo you're working in, on your own computer, and suggests which model to use for it. The scoring code is built in, so it works offline and keeps working even if Agent Friendly Code goes down.

Install

One command works for any supported agent. The vercel-labs/skills CLI finds the agents you have set up and adds SKILL.mdand the scoring code to each one's skills folder.

npx skills add hsnice16/agent-friendly-skill#v0

After installing, run /agent-friendly(or however your agent runs skills) inside any repo on your computer. The skill finds the repo's top folder, scores it, and prints the score with a suggested model. It always gives scores for all 9 agents (Claude Code, Cursor, Devin, GPT-5 Codex, Gemini CLI, Kimi CLI, Aider, OpenHands, Pi). The best fit is picked by score, not by which agent ran the skill, so you get the same result from Claude Code, Cline, Copilot, Continue, or any other supported agent.

How it works

  1. The agent tells you first that it scores the folder you're in, so start from your project's top folder. The CLI also warns you if the folder has none of the usual project files (package.json / README.md / AGENTS.md / .git), so the wrong folder can't quietly give you a low score.
  2. The agent runs node <skill-dir>/dist/index.js .. It's one file (built with ncc) that needs only Node and never uses the network.
  3. It runs the same sixteen checks this site uses (AGENTS.md, CI, tests, README, linter, dev setup, license, contributing guide, pre-commit, dependency list, type config, codebase size, plus four agent-specific instruction files) and gives a score for each agent.
  4. The agent picks the highest score as the best fit (no matter which agent ran the skill) and suggests a type of model using the table below. Switching models is left to you.

Which model for which score

This doesn't favor any provider. The skill suggests a type of model. You pick the exact model for your agent and switch with /model(or your agent's equivalent).

BandScoreSuggested model
High≥ 80Top model: Opus / GPT-5 / Gemini 2.5 Pro. The repo is well set up, so the model can make full use of it.
Mid60 – 79Standard model: Sonnet / GPT-5 Codex / Gemini 2.5 Flash. A good default; a top model is optional.
Low< 60Small, fast model: Haiku / GPT-4o-mini / Gemini 2.5 Flash-Lite. The repo isn't set up well enough to get value from a top model.

Optional: score at the start of every session

If your agent supports session-start hooks, the skill can print a one-line summary at the start of every session. Add one of these to the matching settings file:

Claude Code · .claude/settings.json

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "startup",
        "hooks": [
          {
            "type": "command",
            "command": "node .claude/skills/agent-friendly/dist/index.js . --summary"
          }
        ]
      }
    ]
  }
}

Codex CLI · .codex/hooks.json

{
  "hooks": {
    "SessionStart": [
      {
        "command": "node .agents/skills/agent-friendly/dist/index.js . --summary"
      }
    ]
  }
}

Cursor, Cline, and Copilot don't have a session-start hook yet. Instead, paste the same node ... --summary command into .cursorrules / .clinerules as a fixed instruction, or run /agent-friendly yourself whenever you want a new score.

Works on its own

The scoring code and weights are packed into dist/index.js with @vercel/ncc and saved in the skill repo. Every run only reads files on your computer: no network, no call to this site, no token. If Agent Friendly Code disappears, the skill keeps working. The scoring code is copied from Agent Friendly Code's lib/scoring/(this site's source) and kept in sync by hand, as described in its AGENTS.md.

FAQ

  • Does the skill contact this website?

    No. The scoring code is copied into the skill and packed into its dist/ folder with @vercel/ncc. Once installed, every score runs on your own computer, with no network request. If this site goes offline tomorrow, the skill keeps working the same way.

  • Which agents does it score against?

    Nine: Claude Code, Cursor, Devin, GPT-5 Codex, Gemini CLI, Kimi CLI, Aider, OpenHands, and Pi, the same ones this site scores. It doesn't matter which agent runs the skill: you always get all 9 scores. The same one-line install works in any agent that supports vercel-labs/skills (Cline, Copilot, Continue, Roo Code, Windsurf, Amp, and others).

  • How does it pick a model to suggest?

    After scoring, the skill puts the overall score in a band (high, mid, or low) and suggests a type of model. A high-scoring repo is set up well enough to get real value from a top model (Opus, GPT-5, Gemini 2.5 Pro). A low-scoring repo can't make use of the extra power, so a smaller, faster model is the better choice. The suggestion doesn't favor any provider, and the rules are in SKILL.md.

  • Where is the source?

    github.com/hsnice16/agent-friendly-skill. It's MIT-licensed and uses version tags. To choose a version, add '#<version>' to the install command: `npx skills add hsnice16/agent-friendly-skill#v0` always gets the latest 0.x.y, and '#v0.1.0' stays on one exact version. (The CLI uses '#' for versions, because '@' picks a skill by name.) The scoring code is copied from this site's repo (lib/scoring/) and kept in sync by hand, as this site's AGENTS.md describes.

Source

github.com/hsnice16/agent-friendly-skill — MIT-licensed, with version tags. A sister project to agent-friendly-action . Both use the same scoring code.