Welcome to Understudy Discussions

[s-JoL]

Welcome to Understudy Discussions.

This is the best place for:

• setup and usage questions
• architecture / design discussion
• feature ideas and workflow requests
• sharing demos, skills, and integrations you built with Understudy

A few good starting points:

• README
• Overview
• Product Design
• Discord

If you're interested in contributing, especially valuable areas right now are:

• Linux GUI backend (AT-SPI)
• Windows GUI backend (UIA)
• grounding / validation robustness
• teach-by-demonstration and skill synthesis
• route optimization and memory

Feel free to start a thread, ask a question, or show what you're building.

[s-JoL]

One concrete direction I want to work on next is letting crystallization update an existing skill, not just create a new one.

Right now the main effect of crystallization is publishing a new skill when a repeated workflow becomes clear enough. That works, but over time it can lead to near-duplicate skills even when what we really learned was a better version of something the system already knew.

The rough plan is:

1. Add skill matching before publish
   When a new crystallized workflow is ready, first compare it against existing workspace skills using objective, parameter schema, success criteria, route signature, and key procedure steps.

2. Generate an update candidate instead of always creating a new skill
   If the match is strong enough, synthesize a patch-oriented update candidate rather than a brand new skill.

3. Make the diff human-readable
   The update candidate should explain what changed: new fallback routes, faster routes, tighter verification, broader parameterization, or corrected procedure steps.

4. Apply only safe updates automatically
   Early on, automatic updates should probably be limited to safe cases like adding a better route, adding a fallback, or strengthening validation. Bigger semantic changes should stay conservative.

5. Add versioning / provenance to skills
   Each skill should keep some lightweight history like version, updatedFrom, evidenceCount, and lastValidatedAt so we can track how it evolves.

Conceptually, Layer 3 should not only synthesize new skills. It should also revise and harden existing skills as the system discovers better routes and better recovery patterns.