Phase 7b — Visual Review

Objective

With fresh context, perform a thorough visual review of the built site. Compare screenshots against the enchantment map, evaluate manual criteria 32-34, check image quality, and assess anti-similarity. Produce a structured diagnosis that either approves the site or provides an actionable fix list for the loopback correction agent.

This phase is SEPARATE from Phase 7 (automated scoring). Phase 7 ensures technical quality (score >= 9.0). This phase ensures VISUAL quality and ENCHANTMENT — what makes someone want to buy the site.

Input (provided by orchestrator)

$LEAD_ID — the lead being built
Phase 7 automated scores (from report)
Screenshots at sites/$LEAD_ID/screenshots/
Blueprint at sites/$LEAD_ID/mapa-encantamento.md

Step 1 — Load the 34 criteria

Read: sites/_templates/visual-review-criteria.md

This file contains ALL 34 criteria with IDs, categories, severities, penalties, and pass conditions. You must know every criterion to evaluate properly.

Step 2 — Load the quality rubric

Read: sites/_templates/quality-rubric.md

This contains: criteria 32-34 scoring rubrics, Gate 2 visual analysis questions, visual-analysis.md format template, classic error patterns.

Step 3 — Read the enchantment map

Read: sites/$LEAD_ID/mapa-encantamento.md

Extract and memorize:

Every section's planned LAYOUT, ANIMATION, VISUAL TECHNIQUE
The signature element and where it should appear
The aesthetic direction committed
The animation budget (showstopper, supporting, baseline)

Step 4 — Read Phase 7 automated results

Read: data/reports/site-$LEAD_ID/phase7-quality.md

Note: which automated criteria passed/failed, the final score, any fixes already applied.

Step 5 — Visual inspection of screenshots

Read ALL screenshots with full visual attention:

Read: sites/$LEAD_ID/screenshots/fullpage-desktop.png
Read: sites/$LEAD_ID/screenshots/fullpage-mobile.png
Read: sites/$LEAD_ID/screenshots/viewport-desktop.png
Read: sites/$LEAD_ID/screenshots/viewport-mobile.png

For EACH screenshot, analyze pixel by pixel:

Proportions: Elements with wrong size, distorted images, disproportionate cards
Positioning: Overlapping elements, misaligned, outside container
Spacing: Excessive gaps, inconsistent margins, missing padding
Text: Breaking into too many lines on mobile, overflow, truncation
Responsiveness: Layout working on desktop but broken on mobile (or vice-versa)
Enchantment: Wow moments from the blueprint that were NOT implemented

Step 6 — Evaluate manual criteria 32-34

Criterion 32: Visual Originality

Open screenshots of the last 2 built sites:

LAST_TWO=$(ls -td sites/*/screenshots/fullpage-desktop.png 2>/dev/null | grep -v "$LEAD_ID" | head -2)
for f in $LAST_TWO; do echo "Compare with: $f"; done

Read each and compare. Score 0-10 using the rubric from visual-review-criteria.md.

Criterion 33: Desktop Professionalism

Score from fullpage-desktop.png using rubric.

Criterion 34: Mobile Professionalism

Score from fullpage-mobile.png using rubric. SEPARATE evaluation from desktop.

Step 6.5 — Comparative Framework (MANDATORY before verdicts)

Why this step exists: Without a comparison framework, "Looks like $50K site?" defaults to YES. You need external reference points to honestly evaluate distinctiveness.

6.5a — Load the 5 most recent sites' screenshots

RECENT_SITES=$(ls -td sites/*/screenshots/fullpage-desktop.png 2>/dev/null | grep -v "$LEAD_ID" | head -5)
for f in $RECENT_SITES; do echo "Read: $f"; done

Read each screenshot. Compare structurally against the current site.

6.5b — Fill the Comparison Table (MANDATORY in visual-analysis.md)

For each of the 3 most recent other sites, score similarity 0-10 where 0 = completely different and 10 = identical structure:

## Comparative Framework

### Comparison A — [site name]
- Hero similarity: X/10 — [what specifically is similar or different]
- Navbar similarity: X/10
- Section order similarity: X/10
- Color palette similarity: X/10
- Typography pairing similarity: X/10
- Animation character similarity: X/10
- "Could be confused with" overall: X/10

### Comparison B — [site name]
[same structure]

### Comparison C — [site name]
[same structure]

### Overall Assessment
- **Most dangerous overlap:** [which site + which dimension]
- **Distinguishing features:** List 3 concrete things that make the current site impossible to confuse with any of the 3 comparisons
  1. [Concrete thing #1]
  2. [Concrete thing #2]
  3. [Concrete thing #3]

6.5c — Verdict gate (applies before Step 9 verdicts)

You may ONLY answer SIM/YES to the 3 verdicts in Step 9 if:

ALL 3 "could be confused with" overall scores are <= 5
At least 3 distinguishing features are concrete (not abstract words like "refined", "elegant", "premium")
You can describe the site in 10 concrete words that would NOT describe any of the 3 comparison sites

If any of these conditions fail, the verdicts MUST be NO and you write a Loopback Diagnosis with specific divergence instructions.

Step 7 — Blueprint adherence check

For EACH section in the blueprint, verify:

Was the specified LAYOUT implemented?
Does the specified ANIMATION exist in the component?
Was the VISUAL TECHNIQUE applied?
Is the signature element present where specified?

List every deviation.

Step 8 — Image review

Load the image analysis guide:

Read: sites/_templates/image-analysis-guide.md

Run the image gate:

.claude/scripts/gate-images.sh $LEAD_ID --full

Then perform visual checks from the guide:

Visual inspection in screenshots — overflow, stretch, crop, whitespace
CSS containment in code — object-fit, dimensions, overflow-hidden
Image-context coherence — does image content match section topic?
Contribution — does each image help or hurt?

Step 9 — Answer quality questions + verdicts

Follow the Gate 2 Questions section from quality-rubric.md. Answer ALL questions IN WRITING:

Wow moments, Visual quality, Mobile-specific, Desktop-specific — every question answered.

Three verdicts (answer HONESTLY):

"Looks like a $50K site?" → YES or NO with justification
"Would someone share this link saying 'look how amazing'?" → YES or NO
"If I were the business OWNER, would I be PROUD?" → YES or NO

Step 10 — Produce structured output

If ALL criteria pass (every criterion >= 8, average >= 9.0, all 3 verdicts YES):

Write sites/$LEAD_ID/visual-analysis.md following the format from quality-rubric.md. Mark status as PASS.

If ANY criterion fails:

Write sites/$LEAD_ID/visual-analysis.md with the full analysis AND a Loopback Diagnosis section:

## Loopback Diagnosis

### Status: FAIL
### Loopback Count: {N}/2

### Failing Criteria
| # | Criterion | Score | Required | Category |
|---|-----------|-------|----------|----------|
| {n} | {name} | {score} | >= 8 | {category} |

### Visual Problems Found
1. **[Section: {name}]** {description}
   - File: {file_path}:{line}
   - Fix: {specific fix instruction}

### Blueprint Deviations
- {section}: Blueprint says "{spec}" but implementation uses "{actual}"

### Verdicts
- "Looks like $50K site?" → {YES/NO} — {justification}
- "Would someone share?" → {YES/NO} — {justification}
- "Owner proud?" → {YES/NO} — {justification}

### Suggested Action
{Prioritized list of what to fix first}

CRITICAL: The diagnosis must include specific file paths and line numbers for every problem. The Correction Agent will use this to apply targeted fixes — vague descriptions like "fix the layout" are useless.

Anti-Shortcut Rules

If you find "everything perfect" on first look, you are prioritizing speed. STOP and look again.
visual-analysis.md MUST have >= 1000 characters of real analysis (not checkboxes)
You MUST mention EACH section by name in your analysis
Score 9.0+ from Phase 7 does NOT mean visual review passes — a generic site can score perfectly

Exit Gate

# 1. Visual analysis completeness
.claude/scripts/gate-quality-loop.sh $LEAD_ID --post

# 2. Comparative framework + memorability
.claude/scripts/gate-memorability.sh $LEAD_ID

Both gates must pass. If either fails, return to the corresponding step and fix.

Constraints

Constraint	Enforced by
All 34 criteria >= 8, average >= 9.0	Criteria check
All 3 verdicts = YES	Honest visual assessment
visual-analysis.md >= 1000 chars	gate-quality-loop.sh --post
Every section mentioned by name	Anti-shortcut rule
Diagnosis includes file:line for every problem	Loopback contract
Images coherent and well-contained	gate-images.sh --full + visual check

Exit Criteria

All 34 criteria evaluated (31 automated reviewed + 3 manual scored)
Blueprint adherence checked for every section
Image review complete (gate + visual + coherence)
All quality questions answered in writing
All 3 verdicts answered honestly
visual-analysis.md saved with >= 1000 chars
Status determined: PASS (all clear) or FAIL (with structured diagnosis)
gate-quality-loop.sh --post exits 0

site-phase-7b

Files1

SKILL.md.claude/skills/site-phase-7b/SKILL.md