SkillOpt AGENTS.md

Agent instructions for SkillOpt

Project identity

• Treat this repository as a local-first, open-source SkillOpt fork.
• SkillOpt optimizes skill documents / system prompts, not model weights.
• Prefer examples and smoke tests that work against a local OpenAI-compatible backend.

Default example workflow

When you need a runnable example, default to the bundled DotNetDebug setup:

• config: configs/dotnetdebug/local_mitko.yaml
• sample data: data/dotnetdebug/tasks.json
• seed skill: skillopt/envs/dotnetdebug/skills/initial.md
• output root: outputs/dotnetdebug_smoke

Use small validation runs before suggesting larger experiments. Prefer overrides like:

--cfg-options \
  train.num_epochs=1 \
  train.batch_size=2 \
  gradient.minibatch_size=2 \
  gradient.analyst_workers=1 \
  env.workers=1 \
  env.limit=2

Documentation expectations

• Put the local AI path first in docs and examples.
• Mention cloud backends only as optional alternatives unless the task is explicitly cloud-specific.
• When referring to local inference, prefer openai_compat and http://localhost:8000/v1 unless the user says otherwise.
• Keep the README aligned with .env.example and the shipped configs.

Repo hygiene

• Never commit secrets, .env files, private notes, personal scratch files, local outputs, or benchmark dumps.
• Keep private markdown in ignored files such as *.local.md or *.secret.md.
• Keep IDE state, virtual environments, logs, and generated outputs out of git.
• Before publishing, verify that .gitignore still covers local-only files.

Code and config conventions

• Prefer structured config sections: model, train, gradient, optimizer, env.
• Reuse repo-native configs instead of inventing ad hoc command lines when possible.
• If adding a new backend, update the backend modules, config mapping, runtime dispatch, docs, and env template together.
• If adding a new benchmark, update the env package, config, CLI entry points, and user-facing docs together.