SKILL.md is a much more widely adopted and compatible format for agent harnesses (at least for now). BioSpec naturally falls into this framework, with commands becoming their own skills and templates becoming skill resources. Some things to consider here:
- Instruction quality: Can we improve them based on skill.md best practices
- Instruction depth: Models have become stronger at long-horizon tasks. How can we squeeze more out of a command like 'autofill', which previously hit issues with context overflow and overfocus during repository surveys.
- Granularity: Do any instructions need to be combined into a single skill, or do skills enable new instructions?
- Setup: How should BioSpec templates be initialised in a project?
- Hooks: Should there be hooks to ensure BioSpec skills are triggered if BioSpec has been initialised within a given project/repo?
SKILL.md is a much more widely adopted and compatible format for agent harnesses (at least for now). BioSpec naturally falls into this framework, with commands becoming their own skills and templates becoming skill resources. Some things to consider here: