Jarvis is a goal-driven product-delivery skill suite. Keep it smaller than the workflow systems it replaces.
- Follow
karpathy-guidelines: state material assumptions, prefer the simplest implementation, make surgical changes, and verify observable results. - Follow
efficient-development-workflow: choose checks by risk and avoid mandatory documents, plans, branches, agents, or full suites for routine work. - Explain why guidance exists. Avoid accumulating
MUSTrules to patch isolated failures; add or improve an evaluation scenario instead. - Keep shared policy in
core/and domain lenses incapabilities/. Do not copy them intoSKILL.md. - Treat skill size as a review signal, not a CI gate. Load capabilities, Golden Paths, and Recipes only when the active product or feature needs them.
- Artifacts are optional unless a real downstream consumer needs them.
- Every new behavioral rule needs a failing or discriminating eval scenario.
- Prefer changing a capability or Recipe over broadening the control kernel.
Run before committing:
python scripts/validate.py
python scripts/package_skills.py --check
python -m unittest discover -s tests -vReport skipped checks as skipped. Never claim behavior that the checks did not prove.