eval-banana

Skills

eval-banana ships agent skills in the skills/ directory of the repository. A skill is a bundle of instructions and references that teaches an AI coding agent — Claude Code, Codex, and compatible agents — how to operate a tool. Installing the eval-banana skill lets an agent write, debug, and run checks for you.

Available skills

  • eval-banana — guidance for writing and debugging eval-banana checks: the two check types, auto-discovery rules, the context.json contract, config precedence, and how to read a report.
  • gemini_media_use — helpers for uploading and analyzing image, audio, and video with Gemini, for evals over generated media.
  • voxtral-transcribe — Mistral Voxtral transcription patterns, for evals over generated audio.

Installing skills

Install the repo's skills into your project with the npx skills CLI:

npx skills add https://github.com/writeitai/eval-banana

The CLI auto-detects the AI agents you have installed and copies the skills into their native directories — .claude/skills/, .codex/skills/, .agents/skills/, .gemini/skills/, and so on.

Installed skill directories are installation artifacts: they should usually be gitignored rather than committed into the target project.

What the skill covers

Once installed, the eval-banana skill gives the agent the working knowledge to:

  • choose between deterministic and harness_judge for a given condition;
  • write a valid check file, including the required schema_version, a unique id, and exactly one of script / script_path;
  • read the context.json contract correctly so deterministic scripts locate project files;
  • configure a harness for judge checks and read back the scored report.

It is optional — everything the skill automates, you can do by hand from these docs — but it makes an agent a competent operator of eval-banana inside your repo.