Design Composition Check

Use this before a second feature starts building against a new or unfamiliar design system — after the system exists (tokens, components, or an extracted style guide), before it becomes a shared dependency.

Trigger

Run this when all of the following hold:

• The design system or token set is new or unfamiliar: freshly researched, extracted from a reference, or ingested from a non-technical source — not something already proven across features in this project.
• More than one feature is about to depend on it.
• No prior screen in this project has exercised it yet.

Skip conditions

• An established, framework-driven UI (Material, Ant, shadcn/ui, or similar) already proven elsewhere in this project or organization.
• A routine styling change to an existing, already-validated system.
• No candidate screen in scope scores on any of the three hardness signals (information density, state variety, or primary/critical flow) — state which signal each candidate screen failed before skipping on this ground.

State the skip condition in one line and stop. Do not select screens or render a verdict.

Workflow

1. From the features already in scope, list every screen the design system will need to support.
2. Pick 1-2 of the hardest compositions from that list — never the easiest. See Selecting the hardest compositions for what "hardest" means and how to find it fast.
3. Render each selected screen against the proposed tokens/components: a mockup, a component-driven build, or the actual feature implementation if one is already planned. Let that first real implementation issue double as the evidence instead of requiring a separate throwaway pass.
4. For every place the screen's real content doesn't fit the system cleanly, classify the gap:
   • Correctable — the fix only changes this screen's usage of existing tokens or components (which existing token or variant it reaches for, how it composes them). Nothing in the shared system's definitions changes. Make the fix and note it.
   • Genuine gap — the fix would change a token's definition, add a new token or component, or change an existing component's API or variants — i.e. it changes something every other screen inherits. Do not invent a one-off, screen-local fork. Escalate it as a design-system amendment on the tracking issue.
5. Render one verdict and post it as a comment on the GitHub issue that introduced the design system (or the tracking issue for the feature that will consume it first):
   • Composes cleanly — no gaps found; the system is ready for other features to build on.
   • Composes with a filed gap — every selected anchor screen can be built correctly today even before the gap is fixed; the gap is tracked (per step 4) but nothing blocks shipping now.
   • Does not compose — rework before scaling — at least one selected anchor screen cannot be built correctly until a genuine gap is fixed; the system is blocked until the revision ships.
6. When the verdict is "does not compose," revise the system before any other feature scales on it. Treat the design system as a blocked dependency — pause other features depending on it until the revision ships; do not proceed in parallel hoping the amendment lands later.

Report format

Post this on the tracking issue:

## Trigger
<why this system qualified: new/unfamiliar, more than one feature, no prior screen>

## Screens Selected
- <screen name> — <why it's hard: information density | state variety | primary flow>
- <screen name> — <why it's hard: ...>

## Verdict
<Composes cleanly | Composes with a filed gap | Does not compose — rework before scaling>

## Gaps
- <gap> — Correctable: <what was fixed>
- <gap> — Genuine: escalate as design-system amendment (issue #<n>)

Copy the verdict string exactly as spelled in step 5 above, including the em dash (—, U+2014) in "Does not compose — rework before scaling" — a hyphen or en dash will not match.

Omit the Gaps section only when the verdict is "Composes cleanly."

Guardrails

• Never pick the easiest screens to get a fast "composes cleanly" verdict — that defeats the falsification purpose. If no candidate screen in scope is genuinely hard, use the third skip condition above (and state there which signal each candidate failed) instead of quietly rendering a "composes cleanly" verdict on artificially convenient screens.
• Never let a genuine gap quietly become a one-off, screen-local fork of the system. Escalate it as a design-system amendment on the tracking issue, the same place other features will look before they hit the same gap.
• Do not create a separate manifest, screen-ID registry, or design-pipeline ledger — findings live on the existing GitHub issue.
• This is optional and narrow: it never blocks routine styling work or an already-proven system, and it is not a substitute for to-prd's prototype allowance or the manual acceptance walkthrough in AI_ENGINEERING_WORKFLOW.md — those verify a decision or one feature's behavior; this verifies whether the shared system holds up under its hardest real usages.