c2d2c:d2c — design to code, 1:1
SkillMediac2d2c d2c pipeline — the standard eight-step workflow for implementing or updating UI from Figma designs. Use EVERY time the user provides figma.com/design/... links asking to build, update, restyle, or restore a UI to match a design. Covers survey → plan review → spec extraction → 1:1 build → regression page → motion review → screenshot acceptance → cleanup + MR. Not for exporting code INTO Figma (that is c2d2c:c2d).
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the c2d2c:d2c — design to code, 1:1 skill
What this skill tells your AI
The instructions your AI receives, as published by biasia/c2d2c in skills/d2c/SKILL.md and read by ahel’s review.
Project parameters
Resolve parameters from C2D2C.md at the repo root (fallback: .claude/c2d2c.md). If neither exists, offer to bootstrap one from the template at ../../templates/C2D2C.template.md (plugin root; standalone installs without that file fetch https://raw.githubusercontent.com/BIAsia/c2d2c/main/templates/C2D2C.template.md): fill what the repo itself reveals (token files, gate scripts, regression routes, git host), ask for the rest. This skill uses DESIGN_FILE, TOKEN_DOCS, DS_ROUTE, GATES, MR_TOOL, DEPLOY_ORDER.
Core principles: agree on the plan before building, implement 1:1, every component gets a regression surface, finish with no loose ends. Three points where you MUST stop and wait for user confirmation: step 2 (plan), step 3 (off-system styles), step 7 (acceptance).
1. Skim + survey the current state + reuse audit
- Skim the design with Figma MCP (
get_design_context+ screenshots): structure first, no parameter-digging yet.get_metadatatends to time out on large frames; prefer design context. - In parallel, map the existing implementation: related components, data flow, animation approach (an Explore agent works well).
- Proactively enumerate reuse (don't wait to be asked): for every UI element in the design, look for an existing implementation and produce a reuse matrix (element → component → coupling → reuse strategy):
- What the design shows is usually the design frame of a real component. Assume it already exists, then try to disprove it, not the other way around.
- Coupling tiers: pure presentation (use directly) / light context (does it degrade gracefully without providers?) / heavy data coupling (data layer, runtime, store).
- A
DS_ROUTEpage or demo page rendering the component with mock props is direct evidence the presentation surface is already isolated. - Heavy coupling ≠ no reuse: the default move is to extract the pure presentation core so the live caller, the regression page, and the new surface share it. A visual replica is the last resort, unless the user explicitly wants isolation.
- Produce a diff list: structural differences between design and current state, the heavy-rework areas, and anything worth extracting as shared.
2. Plan review (STOP)
Present the plan for review: scope, component split/abstraction, animation handling, risks. Lead with a recommendation, not a menu. Wait for confirmation before step 3.
3. Extract detailed specs, block by block (conditional STOP)
Read each section with get_design_context + get_variable_defs, down to every gap, font size, and color token:
- Map whatever fits the project's token/component system directly (consult
TOKEN_DOCS). - Report every off-system style to the user, one by one (absorb as a new token? one-off hardcode? change the design?). If everything maps into the system, say so and continue without stopping.
- Figma variable names ≠ code token names. Verify against the project's token files; never guess.
4. Full 1:1 implementation
- Build to the confirmed plan. The target is pixel-level fidelity, not "roughly similar".
- Respect the project gates; UI changes usually touch i18n keys, so handle them in the same pass.
- Run a dev server and compare against screenshots as you go (multi-worktree port contention: read the actual port from your own server log; headless screenshots may render CSS fallbacks instead of WebGL/shaders, verify headed when needed).
5. Regression page (mandatory for component work)
When the change involves a reusable component (new, or visual/state changes to an existing one), create or update its DS_ROUTE/<component> page:
- State enumeration follows the same standard as c2d2c:c2d: every variant axis (read from
tv()/type unions, don't guess), every data shape (value/empty/fallback/truncation/loading/error), every interaction state (rest/hover/focus/selected/disabled + hover popovers). - The page renders from mock props, no live data. It is both the proof that the presentation surface is isolated and the state checklist for a future c2d2c:c2d run.
- For heavily coupled components, extract the pure presentation core first (see step 1) so the live caller and the regression page share it: this kills hand-copied replicas that drift.
- Pure page-level changes (no reusable component) may skip this step; say why.
6. Motion and state review
Walk every animation, transition, and hover/active/loading/empty state in the module:
- Motion tokens compliant? (Projects usually have a motion ladder; inline springs/beziers are banned.)
- Entrances/exits, scroll-linked behavior, reduced-motion fallbacks complete?
- Fix what you find; report the motion inventory and verdict.
7. Screenshot acceptance (STOP)
- What to capture: the regression page, full plus close-ups of key states; if there's no regression page, capture the real route. Use
scripts/ds-shot.mjsbundled next to this SKILL.md (node ds-shot.mjs <route> [--port N] [--dark]) or manual browser capture. - Present them side by side with the Figma
get_screenshotoutput; call out known deviations and why. - Wait for the user to pass acceptance before step 8. Requested changes loop back to step 4.
8. Cleanup + MR
Cleanup (no loose ends, no patches on top):
- Delete every replaced implementation completely: component files, barrel exports, types, i18n source keys, unreferenced styles/tokens, portal mount points, dead constants. Leftover dead code misleads the next investigation.
- Run
GATESto confirm nothing dangles.
MR (runs automatically once acceptance passed and gates are green; do not stop to ask again):
- Rebase onto the latest default branch.
- Open the MR with
MR_TOOL. Description template: Figma node link / acceptance screenshots / change summary / deletion list / gate results /DEPLOY_ORDERdependencies (e.g. "backend first"). - Merge per project convention (e.g. GitLab:
glab mr merge --auto-merge --yeswith server-side rebase; queue auto-merge immediately on fast-moving mains).
Reporting rhythm
- Steps 1-2 merge into one plan-review message (diff list + recommended plan).
- Step 3 stops only for off-system styles; otherwise one line.
- Step 7 is one acceptance message (side-by-side screenshots + deviation notes).
- After step 8, close out: MR link, deletion list, gate results.
Signals
- GitHub stars
- 55
- Forks
- 2
- Last commit
- Aug 2026
Advanced
- Catalog kind
- skill
- Gateway key
d2c- Source
- github.com/biasia/c2d2c