skilltraining-project/skill-training

What it solves

Professionals produce work that is hard to explain as step-by-step instructions. Skill Training lets an agent learn these hidden skills directly from finished human artifacts—without fine-tuning the model, without manually writing rules, and without needing reward signals or success labels. It produces a readable, editable library of skills that measurably improves the agent's performance and transfers to other models.

How it works

The core idea is a loop with four stages:

1. Corrupt – take a professional screenplay and compress it into a bare story outline, removing craft-level detail.
2. Reconstruct – ask the agent to rebuild the full screenplay using a set of skill cards.
3. Compare – contrast the agent's draft with the original human work and describe the gaps.
4. Update – turn that feedback into revised skill cards, each stored as a small Markdown file and committed to a Git history.

The agent's weights stay frozen; only the external skill library changes. Each commit is inspectable, revertible, and reusable. Training requires no new annotations or external reward models.

Who it’s for

Researchers and practitioners who want to improve an agent's domain expertise without GPU-heavy fine-tuning, without hiring domain experts to write skill rules, and without relying on a success signal. It works with closed models, survives model upgrades, and is especially useful when you have a corpus of finished human work but no detailed instructions or labels.

Highlights

• No labels, no rewards – the human artifact itself is the teacher.
• Readable and editable skills – every learned rule is a small Markdown card, and each update is a Git commit.
• Measurable gains that transfer – trained on professional short-drama works, the library improves all six measured writing behaviors on held-out episodes, wins 49 of 75 blind reader judgments, and also helps other models (e.g., Kimi K2.6 and Opus 4.8).
• Runs on commodity hardware – only Python's standard library is needed; no GPU or heavy dependencies.
• Fully auditable – each training run writes generated scripts, loss reports, traces, and the skill library's Git history to a runs/ directory.