Design Language

Capture and enforce a product's visual language — principles + patterns that make it feel like itself. Two modes, one skill.

• Capture — given an external source (Figma URL, screenshot path, live URL), distill novel layout/hierarchy/density/motion decisions and propose a diff to the consumer repo's

  docs/design.md

  .
• Review — given a component or a built-UI screenshot plus an existing

  docs/design.md

  , check fidelity and produce a structured critique that cites specific principles and specific lines.

Advisory by default. This skill never writes to

docs/design.md

Context-efficient design: heavy source extraction (Figma frame parsing, DOM snapshots, screenshot vision) runs in a sub-agent so raw source data stays out of the main thread. Main thread only sees the distilled observations and the proposed diff.

Sub-agent rule: dispatch a sub-agent when extraction involves multiple Figma frames, a full DOM snapshot, or vision analysis of a dense screenshot. Handle directly when the source is a single component file or a short existing diff.

Step 0: Route the Input

Detect from input shape, not a mode flag.

figma.com/design/...

http(s)://...

.png

.jpg

.webp

.tsx

.vue

.svelte

docs/design.md

docs/design.md

docs/design.md

docs/design.md

references/design-md-template.md

If the user's intent is ambiguous (e.g., a screenshot could be inspiration or the product's own UI), ask one clarifying question before routing.

Always load

references/design-heuristics.md

docs/design.md

Mode 1: Capture

Distill a source into proposed additions/changes to

docs/design.md

Step 1: Stage the source artifact

Before extraction, stage a durable copy under

thoughts/design-captures/

YYYY-MM-DD-{short-description}.{ext}

• Figma URL: use

  mcp__plugin_figma_figma__get_screenshot

  to save a PNG of the node, and

  mcp__plugin_figma_figma__get_metadata

  for structure. Save screenshot under

  thoughts/design-captures/

  .
• Live URL: use

  mcp__plugin_chrome-devtools-mcp_chrome-devtools__navigate_page

  +

  take_screenshot

  to capture; save under

  thoughts/design-captures/

  .
• Screenshot path: if outside

  thoughts/design-captures/

  , copy it in with a dated name. If already there, leave it.

The staged artifact is cited in the proposed diff and logged in

docs/design.md §Captures log

Step 2: Extract observations (sub-agent)

Launch a sub-agent to extract observations from the staged source. Keep raw source data (full Figma JSON, DOM tree, pixel analysis) out of the main thread.

Sub-agent prompt:

Extract design observations from this source. Return a structured summary, not raw tool output.

Source type: {figma | live-url | screenshot}
Source location: {URL or staged file path}

Use these tools as relevant:
- Figma: mcp__plugin_figma_figma__get_design_context, get_metadata, get_variable_defs
- Live URL: mcp__plugin_chrome-devtools-mcp_chrome-devtools__navigate_page, take_snapshot,
  evaluate_script (for computed styles of key elements)
- Screenshot: read the image; extract what can be seen.

Reference file: `skills/design-language/references/design-heuristics.md` — use these
section names as your vocabulary.

For each observation:
1. Which heuristic section does it touch? (e.g., §Hierarchy, §Density & Rhythm, §Motion)
2. What specifically is happening? (e.g., "primary CTA uses size + weight + color — three
   levers stacked"; "spacing rhythm tight and consistent between list rows")
3. Can you quantify it? Only if the source is code or Figma; **never** quantify from a
   raw screenshot (vision cannot reliably distinguish 14px from 16px). For screenshots,
   describe qualitatively ("dense", "calm", "airy").
4. Is this observation **novel** to this product, or is it a generic design truth? The
   skill cares only about novel/product-shaping observations.

Return: a numbered list of observations, each with:
- Heuristic section touched
- Specific observation (qualitative or quantitative per rule 3)
- Whether it is novel or generic

Do NOT propose principles or patterns yet. Just observations. Keep under 30 items — if
there are more, return the most distinctive 30.

Step 3: Consolidate before adding (anti-inflation)

Before proposing any new principle or pattern, read the existing

docs/design.md

For each observation returned by the sub-agent:

1. Is there an existing principle or pattern in

   docs/design.md

   that this observation reinforces, refines, or contradicts?
2. If reinforces/refines: propose an edit to the existing entry — tighten the statement, update the citation to include this new source. Do not add a new entry.
3. If contradicts: add it to the Divergences section (see Step 4). Do not silently add a competing principle.
4. If genuinely novel and uncovered: a new entry is justified. State in the proposed diff why no existing principle covers it.

A new principle requires explicit justification that none of the existing N principles cover the observation. Reject the temptation to add when an existing entry could be strengthened.

Step 4: Propose the diff

Output the proposed diff inline. The diff must include:

1. Signature block — see Output Style.
2. Source — what was captured and where it's staged.
3. Edits to existing entries (consolidation wins) — for each, the existing entry and the proposed tightened version.
4. New entries (only if justified) — each with tag (

   [describes]

   /

   [prescribes]

   ), statement, Why, Cite.
5. Divergences — mandatory section, even if empty. What does this source contradict in the current

   docs/design.md

   ? If nothing contradicts, say so explicitly: "No divergences from current principles." Never silently skip.
6. Captures log entry — one line pointing at the staged artifact under

   thoughts/design-captures/

   .

Close with: "Want me to apply this diff to

docs/design.md

Mode 2: Review

Check a component or built-UI screenshot against the product's

docs/design.md

Step 1: Staleness guard

Read

docs/design.md

git log -1 --format=%cd docs/design.md

"

docs/design.md

was last updated {date}. Principles may be stale — a review against stale principles can miss real drift. Continue, or capture fresh first?"

If continuing, note the staleness in the report header.

Step 2: Read the source (code over pixels)

When both code and a screenshot of built UI are available, prefer the code. Claude's vision is unreliable for measurements (14px vs 16px, color stops, precise contrast). Code gives truth.

• Component file path given: Read the file directly. Grep for related tokens (spacing classes, color classes, typography classes). Follow imports one level deep if the component composes from primitives that carry the styling.
• Screenshot only: use vision to observe qualitatively. Bias toward qualitative findings ("dense", "competing primaries", "uneven rhythm") and away from quantitative ("8px grid violated"). Flag that the review is screenshot-only in the report.
• Both: code is source of truth; screenshot is used to cross-check the rendered outcome.

For complex components (multi-file, heavy composition), launch a sub-agent to gather the full surface in one pass:

Read {component path} and any components it composes. Return:
1. All style-bearing declarations (Tailwind classes, style props, CSS imports).
2. Any tokens or design constants referenced.
3. A compact summary of the visual structure (what renders at what level).

Do not interpret. Just gather.

Step 3: Critique against principles, patterns, and heuristics

Walk the source against, in this order:

1. docs/design.md §Principles

   — does the source honor each describes/prescribes entry? For each finding, cite the principle by number.

2. docs/design.md §Anti-principles

   — does the source violate any "we deliberately don't" entries?

3. docs/design.md §Functional patterns

   — if the source is a known functional pattern, does it match the canonical composition? If it's a new pattern, flag that.

4. docs/design.md §Perceptual patterns

   — does the source carry the product's feel (rhythm, motion personality, chrome quietness)?

5. references/design-heuristics.md

   — any heuristic violations not covered above. Cite by section and check ID.

Specificity is required. Every finding must:

• Name the principle or heuristic by its number or ID.
• Cite the exact file:line or element.
• State the violation in one line, not vaguely.
• Propose a concrete fix.

Vague findings ("feels off", "could be tighter", "not quite right") are forbidden. If an observation cannot be tied to a named principle, pattern, or heuristic, drop it — this skill's vocabulary is what it cites. Uncited criticism is slop.

Step 4: Report

Output a structured critique. See Output Style.

If there are no findings, say so plainly: "Review passed. No drift from principles or heuristics observed." — do not pad with generic praise.

The 9 discipline rules (encoded)

These are the skill's guardrails. Cite them back to yourself when drafting output.

1. Anti-inflation. Capture must attempt consolidation before proposing a new principle. A new entry requires explicit justification that no existing principle covers it.
2. Divergences are mandatory. Capture output always has a Divergences section. "No divergences" is a valid answer; silence is not.
3. Code over pixels. Review prefers source code over screenshots when both exist. Capture-from-screenshot biases toward qualitative language.
4. Citations required. Every principle entry in the proposed diff must cite at least one real artifact (code path, staged screenshot, Figma node). No vibes-only principles.
5. Specificity in critique. Review findings cite specific file:line or element and specific principle/heuristic by name. Vague language is forbidden.
6. Descriptive vs prescriptive tagging. Every principle entry is

   [describes]

   (current reality) or

   [prescribes]

   (target). Drift is visible only when the split is explicit.
7. Anti-patterns are first-class. Capture proposes Anti-principles ("we deliberately don't do X") with equal weight to positive principles.
8. Functional vs perceptual patterns stay separate (Kholmatova). Functional = how things work. Perceptual = how things feel. Each in its own section, named by role / quality respectively.
9. Staleness guard. Review checks

   docs/design.md

   mtime; if > 8 weeks old, warn before critiquing.

Every output should be auditable against these rules.

Output style

Always open with a Design View block — mirrors strategic-thinker / product-thinker signature:

`★ Design View ───────────────────────────────────`
- [Mode: Capture / Review / Capture (bootstrap)]
- [Lead observation or finding count]
- [Primary risk: e.g., "2 divergences" / "1 stale principle cited" / "screenshot-only review"]
`─────────────────────────────────────────────────`

Rules:

• Block appears first, before any analysis or diff.
• 2-4 bullets. Assertions, not hedges.
• Dense, direct, peer tone.

Capture output shape

`★ Design View ───────────────────────────────────`
- Mode: Capture ({source type})
- {N} observations distilled, {M} consolidations, {K} new entries, {J} divergences
- {Headline finding}
`─────────────────────────────────────────────────`

## Source
{staged artifact path, date, one-line description}

## Proposed diff to `docs/design.md`

### Edits to existing entries
(…each edit with old → new…)

### New entries
(…each with tag, statement, why, cite — or "None (consolidated into existing)"…)

### Divergences
(…each: this source contradicts principle {N} because … — or "No divergences."…)

### Captures log entry
- `thoughts/design-captures/{filename}` → informed {which entries}

---

Want me to apply this diff to `docs/design.md`?

Review output shape

`★ Design View ───────────────────────────────────`
- Mode: Review ({code / screenshot / code+screenshot})
- {N} findings across {M} principles / heuristics
- {Headline finding or "Review passed"}
`─────────────────────────────────────────────────`

## Scope
- Source: {component path or screenshot}
- `docs/design.md` last updated: {date} {(+ stale warning if applicable)}

## Findings

1. **fails principle {N}: {name}** — {file:line or element}
   - Observation: {one line}
   - Fix: {concrete change}

2. **§{Heuristic section} > {Check ID}** — {file:line or element}
   - Observation: {one line}
   - Fix: {concrete change}

(…or "No findings. Review passed."…)

## Summary
{one-line verdict}

Voice

• Direct and opinionated. "Swap this for a muted gray" not "you might consider a muted gray."
• Peer tone — a design-literate colleague reviewing, not a rubric grader.
• Concise. Every finding earns its place or gets dropped.
• No hedging, no trailing summaries.

Boundaries

• Does not write to

  docs/design.md

  directly. Advisory only; user commits the diff.
• Does not generate new UI from scratch — that is

  frontend-design

  .
• Does not translate a specific Figma frame into production code — that is

  figma-implement-design

  .
• Does not audit accessibility — that is

  a11y-debugging

  . Contrast and focus-ring issues are flagged as design issues (via §Contrast heuristics); WCAG-level audit is out of scope.
• Does not extract or manage tokens — tokens live in code and Figma Variables. This skill references them, does not duplicate them.
• Does not run against multi-product design systems — single-product scope. If a consumer needs to federate a design language across multiple apps, they want a real design-system repo with a component library, not this skill.

Handoffs

When the review or capture reveals a clear next step, offer it.

• Capture mode surfaces a work item (e.g., a

  [prescribes]

  principle that code doesn't match yet) → "Want me to shape this drift into a work item?" →

  /dev-skills:shaping-work

  .
• Review mode surfaces enough findings to warrant a focused fix → "Want me to plan these fixes?" →

  /dev-skills:implementation-planning

  .
• During

  qa-test

  , if design fidelity is in scope → run Review mode alongside functional QA; pass findings forward in the QA report.
• During

  implement-change

  , if a mid-build check is wanted → run Review mode against the in-progress component and feed findings back.

Pass forward: the Design View conclusions, the cited principles/heuristics, and the staged captures — so the next skill doesn't re-read source material.