Shipped something I’ve wanted for a while, and it’s already in the latest beta: PaellaDoc learns from what actually breaks in your projects — not from what a model guesses you need.
What it does
When task runs fail, the system clusters repeated failures that look alike — same shape, same root pattern — and turns each cluster into a pending skill proposal. You review it, and if it’s real, it becomes a skill that handles that pattern next time.
Where the idea comes from
This converges with a direction I’ve been seeing in recent research — e.g. SkillOpt (arXiv:2605.23904), which argues agent skills should evolve through validated edits, not one-shot generation. PaellaDoc takes that one step more conservative: the paper lets a model auto-accept edits when a score improves; PaellaDoc keeps you as the validator. The failures are the signal. You decide what becomes a skill.
Why it’s built this way
Evidence-driven, on purpose. It doesn’t propose a skill because a model “thinks” you need one. It proposes it because the same failure happened, repeatedly, in your real runs. Fingerprints are persisted, so the same pattern is never proposed twice.
Same philosophy as everywhere else in PaellaDoc: the source of truth is what happened, verified — not vibes.
How to use it
- It runs off your task outcomes automatically.
- Proposals show up as a pending action in the Skills view.
- You approve, edit, or dismiss. Nothing is added behind your back.
Local-first. Your failures, your patterns, your machine.
Curious what patterns it surfaces in your projects — drop them here once you’ve run it a while.