gsp-project-build

Phase 5 of the GSP project diamond. Uses a 7-phase pipeline with verification checkpoints to implement designs directly in the codebase as production-ready frontend components.

Works with the dual-diamond architecture: reads brand system from .design/branding/{brand}/patterns/ via brand.ref, reads/writes project assets in .design/projects/{project}/.

Pipeline architecture:

Phase 1: SCAFFOLD (skill-level, no agent)
  └─ /gsp-scaffold → verify build passes

Phase 2: FOUNDATIONS (agent: gsp-project-builder mode:foundations)
  ├─ Context: {brand-name}.yml + token-mapping.md, target-adaptations.md, STACK.md, CONVENTIONS.md
  ├─ Writes: token config, global CSS, layout, shared utils
  └─ CHECKPOINT: build must compile

Phase 3: FOUNDATION REVIEW (interactive)
  └─ Present summary → user confirms

Phase 4: COMPONENTS (agents: gsp-project-builder mode:component, parallel)
  ├─ Orchestrator: reads all design chunks → builds component manifest → partitions
  ├─ Each agent: installs/customizes/creates its assigned components
  ├─ Model assignment: round-robin (Opus/Sonnet) for rate-limit distribution
  └─ CHECKPOINT: build must compile

Phase 5: SCREENS (agents: gsp-project-builder mode:screen, parallel)
  ├─ Context per screen: its design chunk + component paths (components exist in codebase)
  ├─ Agent reads foundations + components from codebase (not from context)
  ├─ Model assignment: round-robin (Opus/Sonnet) for rate-limit distribution
  └─ CHECKPOINT: build must compile

Phase 6: EXTRACTION REVIEW (lightweight)
  └─ Grep for hardcoded values, flag remaining duplication

Phase 7: FINALIZE
  └─ BUILD-LOG, MANIFEST, STATE, phase transition

Implement designs as production-ready code in the codebase via phased pipeline with compile checkpoints.

Input: Design chunks + research chunks + brief chunks + brand system chunks
Output: Code in the codebase + {project}/build/BUILD-LOG.md + {project}/build/SCAFFOLD-LOG.md
Agent: gsp-project-builder (spawned per phase with execution mode)

@${CLAUDE_SKILL_DIR}/../../templates/phases/build.md </execution_context> - Always use `AskUserQuestion` for user interaction — never prompt via plain text - One decision per question — never batch multiple questions in a single message ## Step 0: Resolve project and brand

If .design/projects/ does not exist: output "No GSP project found. Run /gsp-start to begin." and stop.

Resolve project from .design/projects/ (one → use it, multiple → ask). Set PROJECT_PATH.

Read {PROJECT_PATH}/brand.ref → set BRAND_PATH.

Step 0.5: Validate prerequisites

Read {PROJECT_PATH}/STATE.md. Check that Design (Phase 3) is complete or needs-revision (revision means critique ran and is feeding back).
If design is pending or missing: "No designs found. Run /gsp-project-design first — building without designs leads to poor results." Then stop.

Exception: if design_scope is tokens in config.json, skip this check (tokens-only projects don't need design).

Step 1: Load config and check state

Read {PROJECT_PATH}/config.json to get implementation_target, design_scope, codebase_type, app_path.

Set APP_PATH = value of app_path. If empty, default to . (repo root).

Branch check

Read config.json git.branch. If set, check current branch with git branch --show-current. If different, warn: "⚠️ Expected branch {git.branch}, currently on {current}. Switch branches or continue?"

Figma scope check

If implementation_target is figma:

1. Log: "📐 Figma target — producing implementation specs (no codebase to edit)"
2. Skip to Step 7: Figma fallback (single agent, spec-only mode)

Revision mode

Check {PROJECT_PATH}/STATE.md for build status. If status is needs-revision:

1. Read {PROJECT_PATH}/review/issues.md — these are QA issues to address
2. Log: "🔄 Revision mode — addressing QA issues from review/issues.md"
3. Skip to Step 8: Revision mode (single agent with issues)

Design check

If design doesn't exist (no design/ dir or no screen chunks in it), tell the user to run /gsp-project-design first and stop.

Enumerate screens

Read {PROJECT_PATH}/design/ directory. Collect all screen-{NN}-{name}.md files.
Store as ordered list: SCREENS = [(01, landing), (02, changelog-list), ...]

Log screen list for user visibility.

Step 2: Phase 1 — SCAFFOLD

Invoke /gsp-scaffold via the Skill tool.

This handles: dependency installation, config file creation, component library init, build verification.

After scaffold completes, verify {PROJECT_PATH}/build/SCAFFOLD-LOG.md exists. Read it to confirm build status.

Gate: If scaffold reports build failure, stop and surface the error. Do not proceed to foundations with a broken build.

Step 2.5: Load agent methodology

Read ${CLAUDE_SKILL_DIR}/methodology/gsp-project-builder.md. Include the full content as Agent methodology in all agent prompts below (Steps 3, 4.5, 5, 7, 8).

Step 2.6: Load build references

Read these reference files:

• ${CLAUDE_SKILL_DIR}/visual-effects.md
• ${CLAUDE_SKILL_DIR}/../gsp-project-design/block-patterns.md
• ${CLAUDE_SKILL_DIR}/shadcn-composition.md
• ${CLAUDE_SKILL_DIR}/agent-rules.md — common spawn guardrails (universal + per-mode)

Hold their content for inlining into agent prompts in Steps 3, 4.5, 5, 7, and 8.

Step 2.7: Theme apply gate

Verify brand tokens are installed in the codebase before spawning foundations. Foundations no longer pastes tokens — /gsp-brand-apply is the install primitive.

1. If {APP_PATH}/components.json does not exist, skip this gate silently — the target is non-shadcn (scaffold's failure gate at end of Step 2 already covers a broken shadcn scaffold).
2. Resolve the CSS path: read {APP_PATH}/components.json and extract the value at .tailwind.css (a relative path from APP_PATH).
3. Open {APP_PATH}/{cssPath}. Verify:
   • Contains oklch(
   • Has both :root { and .dark { blocks
   • Declares --background, --foreground, --primary, --radius
   • If {BRAND_PATH}/patterns/{brand-name}.theme.json exists: contains the brand's signature cssVars.light.background value from that file (this distinguishes the applied brand from shadcn's nova defaults). If {brand-name}.theme.json does NOT exist (older brand from before Task 4 landed), skip the brand-signature check and rely on the structural checks only — log ⚠ Brand theme.json not found — skipping brand-signature check.

If any required check fails, brand tokens are not applied (or the wrong brand is applied). Use AskUserQuestion:

• Question: "Brand tokens for {brand-name} not detected in {APP_PATH}/{cssPath}. Run /gsp-brand-apply {brand-name} now?"
• Options:
  • A: "Yes — I'll run /gsp-brand-apply {brand-name}, then re-run /gsp-project-build"
  • B: "No, abort the build"

On A: output Next: run /gsp-brand-apply {brand-name}, then re-invoke /gsp-project-build and exit this skill. The build does not auto-continue — the apply runs out-of-band and you re-invoke when ready.

On B: stop the build phase. Output: Build aborted — apply brand tokens with /gsp-brand-apply {brand-name} and re-run /gsp-project-build.

If all checks pass, log ✓ Brand tokens verified and continue to Step 3.

Step 3: Phase 2 — FOUNDATIONS

Spawn gsp-project-builder agent with execution_mode: foundations.

Context for foundations agent (lean — no screen chunks):

File	Purpose
{BRAND_PATH}/patterns/{brand-name}.yml	Token values only — used with token-mapping.md to generate CSS variables. Do NOT re-read patterns/constraints/effects from here — those are in STYLE.md.
{BRAND_PATH}/patterns/STYLE.md	Design law — philosophy, patterns, constraints, effects, bold bets, implementation hints (if exists; fall back to {brand-name}.md)
{PROJECT_PATH}/brief/target-adaptations.md	Component adaptations for target
.design/system/STACK.md	Stack state (or .design/system/stacks/{APP_NAME}.md for monorepos)
.design/system/CONVENTIONS.md	Codebase conventions (if exists)
.design/system/COMPONENTS.md	Existing components (if exists)
{PROJECT_PATH}/config.json	Tech stack, target, app_path
APP_PATH = {APP_PATH}	Working directory — all file writes and build commands run here
Build output template (from execution_context)	Build log structure
Token mapping ref (loaded in Step 2.6)	shadcn component composition rules, semantic token usage, cn(), cva, RSC patterns
Visual effects, block patterns refs (loaded in Step 2.6)	Design patterns + CSS recipes
Agent methodology (loaded in Step 2.5)	Builder role, process, quality standards

Agent instructions:

execution_mode: foundations
Build token integration, global styles, layout primitives, theme provider/utilities ONLY.

Foundation-specific actions:

1. Add base styles + dark mode setup + font imports that apply did not handle (cssVars.theme.font-sans may set the CSS var without the next/font/google import — add it if missing)
2. Create root layout with nav + footer shells (structure only)
3. Create shared utilities (cn helper, theme provider)
4. Apply STYLE.md bold bets + effects as CSS utilities / Tailwind extensions; validate against never/always constraints
5. For shadcn targets: semantic tokens (bg-primary), never raw values (bg-blue-500); gap-* not space-y-*; size-* when w/h equal

See agent-rules.md (loaded in Step 2.6) for universal + per-mode guardrails (write to codebase, leave unstaged, log to build/logs/foundations.md, no BUILD-LOG.md direct write).

Checkpoint: Compile check

After the foundations agent completes, run the build command in APP_PATH:

Stack	Build command
Next.js	cd {APP_PATH} && npx next build
Vite	cd {APP_PATH} && npx vite build
TypeScript only	cd {APP_PATH} && npx tsc --noEmit
Generic	cd {APP_PATH} && npm run build

Pass: Continue to preview verification, then Step 4.
Fail: Log the error. Do NOT re-spawn the agent. Surface the error to the user and ask how to proceed.

Preview verification (opt-in)

After compile passes, verify the foundations actually render:

1. Check if dev server is already running (lsof -i :3000 or :5173)
2. If running, use curl -s http://localhost:{port} to fetch the page
3. Check the HTML response for:
   • Not blank — response body has more than just the shell/boilerplate (>500 chars of content)
   • Tokens resolved — grep the response for CSS variables or Tailwind classes from the token config. If var(-- appears but no matching custom property is defined, tokens may be broken.
   • Font loaded — check for the expected Google Fonts import or @font-face rule

If dev server is not running, skip verification silently — do not start one. This keeps it zero-config.

Report any issues found: "⚠️ Preview check: {issue}. This may be cosmetic — continue or investigate?"

Step 4: Phase 3 — FOUNDATION REVIEW

Present a summary of what the foundations phase produced:

  ◆ foundations complete

    Files created/modified:
    - {list from BUILD-LOG.md}

    Tokens: {integrated / skipped}
    Layout: {created / modified}
    Build: compiles ✓

  ──────────────────────────────

Use AskUserQuestion: "Foundations look good? Continue building components, or review first?"

• Continue → proceed to Step 4.5
• Review first → pause, let user inspect, resume when ready
• Adjust → user requests changes (colors, typography, spacing, etc.)

Brand feedback loop

Read ${CLAUDE_SKILL_DIR}/brand-feedback.md for the full procedure. Key constraint: the gsp-brand-engineer re-sync runs synchronously before Step 4.5 begins.

Step 4.5: Phase 4 — COMPONENTS

Build component manifest + classify + partition

Read ${CLAUDE_SKILL_DIR}/component-classification.md for the manifest schema, classification table (library-default / library-customize / custom / existing), and partitioning rules (≤5 → single agent; group related variants; 3-6 components per agent).

Resume check

Check for existing build/status/component-*.json files. For each partition with a "status": "complete" file, skip that agent — log: "Skipping {name} — already complete."

Progress log

Before spawning, log the manifest:

  ◆ components phase

    Spawning {N} agents in parallel:
    {for each partition}: [{model}] {partition-name} — {component-count} components

Spawn component agents in parallel

For each partition, spawn gsp-project-builder with execution_mode: component.

Assign models in round-robin: first agent on user's model, second on sonnet, third on user's model, etc. This splits rate-limit pressure across model buckets.

Context per component agent:

File	Purpose
Component partition (list + classifications + overrides)	What to build
{BRAND_PATH}/patterns/STYLE.md (or fallback {brand-name}.md)	Design constraints, effects vocabulary
{BRAND_PATH}/patterns/{brand-name}.yml	Token values
{BRAND_PATH}/patterns/components/token-mapping.md	Component-to-token mapping
Design chunk excerpts (only sections referencing these components)	Usage context — how screens use them
{PROJECT_PATH}/brief/target-adaptations.md	Component adaptations for target
{PROJECT_PATH}/config.json	Tech stack, implementation target
Visual effects, block patterns refs (loaded in Step 2.6)	Design patterns + CSS recipes
Agent methodology (loaded in Step 2.5)	Builder role, process, quality standards

Agent instructions template:

execution_mode: component
implementation_target: {target}
components: [{partition list with classifications}]

Install, customize, or create the assigned components per their classification (library-default → install as-is; library-customize → install + STYLE.md overrides; custom → create from scratch).

See agent-rules.md (loaded in Step 2.6) for guardrails (read foundations from codebase, do not modify foundation files, log to build/logs/component-{partition-name}.md, write build/status/component-{partition-name}.json for resume support).

Checkpoint: Compile check

After ALL component agents complete, run the build command (same stack table as Step 3 checkpoint).

Pass: Continue to Step 5.
Fail: Log the error. Surface to user: "Component build failed: {error}. Fix now or skip to screens?"

Merge component logs

After the compile checkpoint passes, merge all build/logs/component-*.md files into {PROJECT_PATH}/build/BUILD-LOG.md (foundations section from build/logs/foundations.md + all component sections, in partition order).

Log: " ✓ components complete — {N} agents, build compiles"

Update {PROJECT_PATH}/STATE.md — set completed component partitions in build status.

Step 5: Phase 5 — SCREENS (parallel)

Build all screens in parallel. Components exist in the codebase from Phase 4.

Context per screen (lean — only this screen's data):

File	Purpose
{PROJECT_PATH}/design/screen-{NN}-{name}.md	This screen's design chunk
Component file paths from BUILD-LOG.md components section	Where to import from (paths only — agent reads codebase)
{PROJECT_PATH}/brief/target-adaptations.md	Component adaptations
{PROJECT_PATH}/research/reference-specs.md (if exists)	Technical specs — include only sections relevant to this screen
{PROJECT_PATH}/critique/prioritized-fixes.md (if exists)	Critique fixes — include only fixes tagged to this screen
Build output template (from execution_context)	Build log structure
Visual effects, block patterns refs (loaded in Step 2.6)	Design patterns + CSS recipes
Agent methodology (loaded in Step 2.5)	Builder role, process, quality standards

Does NOT receive: other screen chunks, brand .yml (already in codebase), full brand system, research monoliths, component source code (agent reads from codebase).

Resume check

Check for existing build/status/screen-*.json files. For each screen with a "status": "complete" file, skip that agent — log: "Skipping screen-{NN}-{name} — already complete."

Progress log

Before spawning, log:

  ◆ screens phase

    Spawning {N} agents in parallel:
    {for each screen}: [{model}] screen-{NN}-{name}

Spawn screen agents in parallel

For each screen in SCREENS, spawn gsp-project-builder with execution_mode: screen.

Assign models in round-robin: first screen on user's model, second on sonnet, third on user's model, etc.

Agent instructions per screen:

execution_mode: screen
screen: {name} ({NN})

Build the {name} screen. Foundations + components are already in the codebase.

1. Read existing layout, tokens, utilities, components from the codebase
2. Create the route page + screen-specific components
3. Wire imports to existing foundation + component files
4. The brand's visual effects exist as utilities/classes from foundations — use them, don't redefine

See agent-rules.md (loaded in Step 2.6) for guardrails (no modifying foundations or shared components, log to build/logs/screen-{NN}-{name}.md, write build/status/screen-{NN}-{name}.json for resume support).

Checkpoint: Compile check

After ALL screen agents complete, run the build command (same stack table as Step 3 checkpoint).

Pass: Log success, continue to Step 5.5.
Fail: Log the errors. Present to user: "Build errors after screens phase: {errors}. The following screens may have issues: {list}. Fix now or continue to extraction review?"

Merge screen logs

After the compile checkpoint passes, merge all build/logs/screen-*.md files into {PROJECT_PATH}/build/BUILD-LOG.md (append screen sections in order: 01, 02, 03, etc.).

Log: " ✓ screens complete — {N} screens, build compiles"

Update {PROJECT_PATH}/STATE.md ## Screen Build Status table — set completed screens to complete.

Step 5.5: Extraction review (lightweight)

Read ${CLAUDE_SKILL_DIR}/extraction-review.md for the full procedure. Quick post-build sanity check: hardcoded values + duplicated patterns. If issues found, AskUserQuestion (Fix inline / Continue to Step 6).

Step 6: Finalize

After all screens complete (or pipeline stops):

Write INDEX.md

Write {PROJECT_PATH}/build/INDEX.md:

# Build
> Phase: build | Project: {name} | Generated: {DATE}

| Chunk | File | ~Lines |
|-------|------|--------|
| Scaffold Log | [SCAFFOLD-LOG.md](./SCAFFOLD-LOG.md) | ~{N} |
| Build Log | [BUILD-LOG.md](./BUILD-LOG.md) | ~{N} |

Write manifest

Write {PROJECT_PATH}/codebase/MANIFEST.md from templates/manifest.md:

1. Components table — one row per component produced. Action = added or modified based on .design/system/COMPONENTS.md. File paths reference actual codebase locations.
2. Patterns table — patterns established (infer from BUILD-LOG.md).
3. Files Touched — flat list of all codebase file paths from BUILD-LOG.md.

Update exports index

Update {PROJECT_PATH}/exports/INDEX.md — add build phase entries between <!-- BEGIN:build --> and <!-- END:build --> markers. Reference build/BUILD-LOG.md and build/SCAFFOLD-LOG.md.

Update state

Update {PROJECT_PATH}/STATE.md:

• Set Phase 5 (Build) status to complete (if all screens done) or in-progress (if partial build)
• Record completion date
• Update ## Screen Build Status table — set Build Status per screen (complete/partial/pending)

Phase transition output

Invoke /gsp-phase-transition with phase build and output directory {PROJECT_PATH}/build/.

---

Step 7: Figma fallback

Read ${CLAUDE_SKILL_DIR}/flows/figma.md for full instructions.

For implementation_target: figma, skip the phased pipeline. Produce Figma-ready implementation specs instead of editing the codebase. Then continue from Step 6 (finalize).

Step 8: Revision mode

Read ${CLAUDE_SKILL_DIR}/flows/revision.md for full instructions.

For needs-revision status, fix QA issues from review/issues.md via a single revision agent. Then continue from Step 6 (finalize).