Load Order
Before the first Read of this skill: Read mstar-harness-core (SKILL.md) — lifecycle, authorization, and conflict authority. The design contract, when the baseline is a written contract, is owned by mstar-design-md; this skill assumes an authoritative baseline exists and defers that contract's authority and sync direction to it; this skill does not restate those.
| Also Read | When |
|---|---|
mstar-design-md | any difference touches tokens, completeness levels, light/dark, or the baseline-vs-document authority rule |
mstar-coding-behavior | the round ends in code changes rather than document changes |
On conflict, mstar-harness-core wins; on a design-contract question, mstar-design-md wins.
Scope
In scope: an authority already exists and an implementation has to be converged onto it — locating and making the baseline readable, deriving the machine-readable fixture, measuring parity per layer, triaging each difference, closing it in one direction, and reporting what was actually compared.
| Out of scope | Where it belongs |
|---|---|
| Designing a system where no baseline exists | not this skill — decide the contract first |
DESIGN.md sections, token naming, completeness levels, light/dark rules | mstar-design-md |
| Prototype-vs-document authority and its one-directional sync | mstar-design-md § Prototype vs document — read it there; this skill only applies it |
| The comparators, CI jobs, and generators of a specific repository | that repository's own checks |
Workflow
- Baseline — name the authority: its kind, its location, whether it is readable and tracked, and what was redacted before it was committed. No baseline → stop, this skill does not apply.
- Machine — derive what can be derived instead of typing it: fixture, snapshot, extraction. Record the regeneration command and its check mode in the same place.
- Parity — compare per layer: structure, appearance, behaviour. Every finding carries the count it was measured on; a probe that matched nothing says so instead of reporting zero defects.
- Triage — classify every difference as defect, decision, or phantom. A decision is written into the contract record with its reason, never silently into the implementation.
- Close — change the implementation, or amend the baseline when triage recorded a decision.
- Evidence — return what
## Evidencebelow describes; a verdict without its counts is not one.
Decision Rules
- A baseline that is not in the repository is not a baseline. Tracked, redacted, and read by the comparators. A baseline in a chat transcript, a screenshot, or a gitignored working directory cannot be converged onto — convert it into a tracked file first, or state that the round has no baseline.
- A baseline that no CI job reads is also not a baseline. Committing closes one of three links: committed → every comparator points at it and strict mode is the default → a CI job actually calls it. Never declare a check live before all three close; a fixture only a human runs is documentation.
- Generated artifacts are never hand-edited. Which artifact wins and which direction the sync runs is
mstar-design-md's rule — this skill applies it, it does not restate it. - Do not invent the baseline. No rounding a source value, no adding a component the source does not define, no drawing an asset it does not ship, no inference from an unreadable source. Silence in the baseline is a decision, recorded with its reason.
- A check that cannot say what it covered did not pass. Empty probe ≠ zero overflow. Every verdict names the count it compared — including the failure path: "baseline missing" must also state how much was owed, or it reads as a pass at exit 0.
Evidence
A finished round returns three things:
- The classified list — every difference labelled defect / decision / phantom, with the unit it belongs to; a decision also names where it was recorded.
- The coverage statement — counts compared per layer, plus what was skipped and why. "Everything matches" without a count is not a coverage statement.
- The strict-mode run — the comparator invoked with failure fatal, output pasted as run, not summarized.
A screenshot, a rendered page, or a green check with no count is not evidence for any of the three.
References
| When | Open |
|---|---|
| choosing a baseline kind, making it machine-readable, closing the three-link chain, redacting before committing | references/baseline.md |
comparing structure / appearance / behaviour, choosing a measurement, the @media vs @container trap | references/parity.md |
| classifying a difference, recording a decision, checking a suspected finding against the phantom catalogue | references/triage.md |
| the normative design contract, token rules, completeness levels, baseline-vs-document authority | mstar-design-md |
