-
Notifications
You must be signed in to change notification settings - Fork 0
Design Principles
These govern every mode of the skill. They come from post-mortems of real
harness failures, so they're treated as load-bearing. The canonical statement
lives in SKILL.md;
this page is the readable version.
A rule that lives only in prose will eventually be ignored. Each rule should sit as high on this ladder as the project tier allows:
prose → checklist → template → hook with exit code → tested code
When you install a prose rule, note what its executed backstop would be.
Plans, unknowns, decisions, statuses — all files. Any session must be able to resume from disk alone.
No transition without a validated artifact: an executed check's output, or a file that exists and is non-empty. An LLM-emitted number or claim is never evidence.
A guard that has never been proven to fire must be assumed hollow. The harness self-test feeds each guard a fake violation and requires rejection.
When auditing, read the actual files and run the actual commands. A doc claiming "X is enforced" is a claim to check, not a fact.
Over-building the harness is the same scope-creep disease it exists to prevent. Install the smallest Tiers that fits; add machinery only when a retro shows the same failure twice, and record what incident justified each addition. Every mechanism should be able to answer "what failure created you?"
Reference by ID or pointer; never copy text between docs. A copy is a fork that will rot — and the The Phase Audit actively hunts for copied-text forks.
In one breath: prompting is not enforcing; state lives on disk; status is derived from evidence; every guard ships a plant test; audits verify disk reality; right-size then grow on felt pain; one source of truth per fact.
Concepts