realdealcpa-vr/vr-agent/skill-tuner

Skill Tuner — optimize for the chosen LLM

Different models need different handling. A frontier reasoner does best with lean, principle-first skills and long autonomy; a small or local model needs explicit micro-steps, tighter loops, and lower reasoning effort. This skill offers that as an option — it never changes anything until asked.

When to use

• The partner switched the model (vragent.ps1 model) or said which model they're running.
• They ask to "optimize/tune for <model>", "I'm on a smaller/local model", or "why is it slow/over-thinking".

The tool

scripts/optimize-skills.ps1 reads vr-overlay/config/model-profiles.yaml, matches the active model to a tier (frontier / strong / mid / small), and applies that tier's settings. Run it from the repo root.

What each tier changes (config)

• reasoning_effort — high for frontier/strong, medium for mid, low for small.
• max_turns — longer autonomy for stronger models, tighter loops for smaller ones.
• skill_style (deep mode) — lean / scaffolded / explicit guidance density.

How to run it as the agent

1. Get the active model: read model.default in the config (or what the partner names).
2. Run the offer first (no -Apply) and show the partner what it would change.
3. If they approve, run with -Apply (add -Deep only if they want the skill text re-tuned

for a notably weaker/smaller model).

1. Confirm with vragent.ps1 doctor and a quick skills list.

Guardrails

• Default is offer-only; never apply tuning without the partner's go-ahead or an explicit -Apply.
• The deep re-tune writes overrides, never touching the tracked canonical skills — -Revert undoes it.
• Tuning changes how skills are delivered, never the authority gates or the accounting correctness.