ptgp ships a skill file for VFE (.claude/skills/ptgp-vfe/) that describes when to use VFE, how to diagnose training issues, common failure modes, and tips for getting good results. It would be good to have similar skills for other GP methods (SVGP, exact GP, VFF, spherical harmonics, etc.).
These skills capture the kind of knowledge that's hard to put in a docstring: when to use which approximation, what to check when training goes wrong, folk wisdom about initialization and hyperparameter choices.
Another area where skills could help is setting priors on kernel hyperparameters. Choosing priors for lengthscale, amplitude, and noise is one of the trickiest parts of GP modeling for practitioners, and the right choices depend heavily on the data and the problem. A skill that encodes guidance on prior selection (e.g. how to think about lengthscale relative to input range, when to use informative vs. weakly informative priors on sigma) would be valuable.
One possibility: skills could be used to help generate "choose your own adventure" style documentation. With skills handling the interactive, problem-specific guidance, documentation wouldn't need to be exhaustive and prescriptive. Instead it could be short, succinct case studies that don't have to go into the weeds of debugging or the use of every utility and helper. The skills files would be especially useful for users working with an LLM to plan and discuss how to solve their particular modeling problem.
Open questions
- How should skills be installed for users? Currently they live in the repo under
.claude/skills/. What's the right distribution mechanism?
- What methods should have their own skill file?
- What should the skills cover beyond what the VFE skill already does?
ptgp ships a skill file for VFE (
.claude/skills/ptgp-vfe/) that describes when to use VFE, how to diagnose training issues, common failure modes, and tips for getting good results. It would be good to have similar skills for other GP methods (SVGP, exact GP, VFF, spherical harmonics, etc.).These skills capture the kind of knowledge that's hard to put in a docstring: when to use which approximation, what to check when training goes wrong, folk wisdom about initialization and hyperparameter choices.
Another area where skills could help is setting priors on kernel hyperparameters. Choosing priors for lengthscale, amplitude, and noise is one of the trickiest parts of GP modeling for practitioners, and the right choices depend heavily on the data and the problem. A skill that encodes guidance on prior selection (e.g. how to think about lengthscale relative to input range, when to use informative vs. weakly informative priors on sigma) would be valuable.
One possibility: skills could be used to help generate "choose your own adventure" style documentation. With skills handling the interactive, problem-specific guidance, documentation wouldn't need to be exhaustive and prescriptive. Instead it could be short, succinct case studies that don't have to go into the weeds of debugging or the use of every utility and helper. The skills files would be especially useful for users working with an LLM to plan and discuss how to solve their particular modeling problem.
Open questions
.claude/skills/. What's the right distribution mechanism?