Auction Item Image Analysis Reference

Use this skill when working on image analysis features, Claude Vision prompts, or photo-based cataloging. Focus on: $ARGUMENTS

How Image Analysis Works in the Extension

Architecture

Item images (up to 5 slots) → base64 encoding → Claude Vision API (Sonnet 4.5)
                                                      ↓
                                               Structured JSON response:
                                               - Title, description, condition
                                               - Artist, period, materials
                                               - Keywords, estimate, reserve
                                               - Image quality assessment
                                               - Sure Score confidence

Model Selection

Sonnet 4.5 — all image analysis (single and multi-image). Uses apiManager.getCurrentModel().id which defaults to Sonnet.
Opus 4.6 — only for valuation request clustering (_clusterImages()) and valuation generation, with Sonnet fallback on overload.
Haiku is NOT used for image analysis.

Image Slots (categorized)

Slot	Purpose	What to extract
Front/Main	Primary identification	Object type, style, material, overall condition
Back/Reverse	Maker marks, construction	Stamps, signatures, labels, construction method
Markings	Stamps, hallmarks	Manufacturer, date, origin, purity (silver/gold)
Signatures	Artist identification	Name, date, numbering, technique notes
Condition	Damage documentation	Specific flaws, repairs, wear patterns

Image Processing

Anthropic API limit: 5MB base64 string; threshold set at 4.5MB for safety margin
File upload limit: 10MB (maxFileSize in config)
Resized via canvas using ensureBase64WithinLimit() in ai-image-analyzer.js
Also _resizeBase64Image() in valuation-request-assistant.js (separate implementation, with MAX_ATTEMPTS safety guard)
Max dimension: 1400px, progressively reduced by 200px if still too large
Output: JPEG at 0.82 quality (minimum 0.5)
Domain whitelist: auctionet.com, images.auctionet.com, upload.wikimedia.org

Two Analysis Paths

Single image (analyzeImage): 45s timeout, 1200 max_tokens
Multiple images (analyzeMultipleImages): 90s timeout, 1500 max_tokens, minimum 2 images required
System prompt sourced dynamically from getAIRulesManager().getSystemPrompt('freetextParser') with fallback chain: freetextParser → core → minimal hardcoded prompt

Sure Score Confidence System

calculateSureScore() provides a weighted composite confidence:

Image analysis quality (40%)
Image quality score (25%) — clarity, lighting, angle, completeness
Object identification confidence (20%)
Market validation (15%) — cross-references with Auctionet historical sales

Levels: "Mycket låg", "Låg", "Medel", "Hög", "Mycket hög"

Market Validation Integration

validateWithMarketData() cross-references image analysis with Auctionet sales data. Conservative scaling with multipliers 0.4–1.0 based on market support percentage. Builds search queries automatically from image analysis output.

Image Clustering (Valuation Requests)

_clusterImages() groups multiple customer images by object for multi-item valuation. Uses Opus 4.6 with Sonnet fallback. Cataloger can drag/drop to adjust groupings.

Multi-Image Upload UI

generateMultipleImageUploadUI() creates a drag-and-drop upload grid with slot-specific zones (front, back, markings, signature, condition).

What to Look For (by category)

Art & Paintings

Signature: location (lower right, a tergo), legibility, full name vs initials
Dating: exact year, decade, century — write exactly as on work
Technique: oil on canvas, watercolor, mixed media, lithograph, etching
Numbering: edition number for prints (e.g., "120/310")
Frame: type, period (but NEVER report frame condition)
Canvas/paper: condition, foxing, tears, restorations
Size: height × width without frame

Furniture

Style identification: gustavian, rococo, art deco, functionalist
Construction: joinery type, original vs restored parts
Wood type: visible in unfinished areas — only identify if confident
Hardware: original vs replacement handles, locks, hinges
Surface: veneer condition, inlay, painted decoration
Stamps/labels: maker marks, often on underside or back

Silver & Gold

Hallmarks: city mark, year mark, maker mark, assay mark
Purity: sterling (925), 830, 18k, 14k — only from stamps, never guess
Weight: essential for precious metals
Engravings: monograms, presentation inscriptions (always mention in condition)
Construction: filled base vs solid (affects weight relevance)

Glass & Ceramics

Manufacturer marks: underside stamps, etched signatures
Technique: blown, pressed, cut, etched, overlay
Color: describe accurately for glass
Condition: chips (nagg), hairline cracks (hårspricka), crazing
Size: height and/or diameter

Rugs & Textiles

Origin: Persian, Turkish, Scandinavian — from patterns and construction
Construction: hand-knotted, machine-made, flatweave
Material: wool, silk, cotton — from texture and sheen
Condition: fringe wear, color fading, repairs, moth damage
Size: always in title — length × width

Watches & Clocks

Manufacturer & model: from dial and caseback
Movement: automatic vs quartz (from caseback or movement view)
Case material: steel, gold, gold-plated — from stamps
Case size: diameter in mm
Dial condition: patina, lume condition, text legibility

Weapons & Militaria

EXTRA CAUTION — anti-hallucination rules are strictest here
Only describe what is physically visible
Never add historical context not explicitly present
Never guess school names, regions, or periods from style
Focus on: markings, inscriptions, materials, construction method

AI Prompt Patterns for Image Analysis

Hallucination Prevention

"Describe ONLY what you can see in the image"
"Do NOT infer historical period from style alone"
"If uncertain about material, say 'appears to be' not 'is'"
"Never fabricate maker names, dates, or provenance"

Extraction Priority

Object identification — what is it?
Material identification — what's it made of?
Maker/artist identification — who made it? (from visible marks only)
Condition assessment — what damage/wear is visible?
Period estimation — when was it made? (conservative)
Valuation clues — quality level, rarity indicators

Multi-Image Strategy

When multiple images are available:

Cross-reference front view with detail shots
Use back/underside for construction clues
Use marking photos to confirm or identify maker
Use condition photos to quantify damage described in main view
Don't repeat information already captured from another image

Photography Guidelines (for reference)

Minimum Standards

At least 3 photos per item
Clean, neutral background (NCS S4000 N gray)
Even lighting, no harsh shadows
Include scale reference where helpful

Category-Specific Photo Requirements

Category	Required shots
Furniture	Front, back (if old), marble top detail, interior
Art	Without frame (expensive), with frame, signature, back
Silver/Gold	Full view, hallmarks close-up, engravings
Porcelain	Full view, underside/marks, pattern detail
Rugs	Full spread, corner detail, back
Watches	Dial, caseback, movement (if accessible)
Services	Group shot, single piece detail, marks

image-analysis

Files1

SKILL.md.claude/skills/image-analysis/SKILL.md