best-practices: a six-axiom kernel for the disciplined program.md author

[AgriciDaniel]

posting under "show and tell" because the README invited it, and because there is a fit to the program.md pattern that may save someone reinventing the wheel.

shipped a small repo:

https://github.com/AgriciDaniel/best-practices

it is a kernel for shipping changes (with humans or AI agents). six axioms in three acts:

before read · name
during small · delete
after evidence · failure

three breaths: read first. write second. verify third.

an agent kernel layers on top: one chair, bounded slices, explorer / worker / verifier roles, five-part closeout. autoresearch is a clean instance of this shape already: the agent edits one file (train.py), under one chair (the human authoring program.md), inside a fixed time budget (the slice), measured by one metric (val_bpb is the acceptance criterion). bounded scope, evidence over intuition, undo plan as kept-or-discarded.

posting in case the kernel is useful as a reference for anyone writing more elaborate program.md files, or composing autoresearch-style swarms.

framing: this is a meditation, not enforcement. for adversarial-agent guards i compose with obra/superpowers in the README. different altitudes.

MIT.