To specify a system that lets AI intelligently select and "drag-crop" chat bubble ranges (instead of raw pixel cropping), define these four pillars:

1. Detection stack: Object Detection (bubble bounding boxes) + OCR (text content) + Contextual Anchor Logic (evidentiary keywords)
2. Decision flow: Detect → Read → Anchor (start/end) → Compute crop region → Output
3. Conditions: boundary contrast, keyword dictionary, aspect-ratio preservation
4. Prohibitions: never bisect a bubble, never crop without context, never strip timestamps/profile marks

Use this skill to produce a full system spec (methods, step-by-step flow, conditions, prohibitions) in the same structure every time, adaptable to any chat platform or evidence domain (not just financial fraud — also harassment, contract disputes, etc.).

Progress:

• Step 1: Define the object taxonomy for the target chat UI
• Step 2: Specify the OCR/text-analysis layer
• Step 3: Define the anchor keyword dictionary for the domain
• Step 4: Write the 5-step decision flow (Detect → Read → Anchor → Compute → Crop)
• Step 5: Enumerate system conditions (contrast, keyword set, aspect ratio, output canvas)
• Step 6: Enumerate strict prohibitions (no bisecting, no floating context-free bubbles, no metadata stripping)
• Step 7: Validate spec against edge cases (oversized bubbles, ambiguous anchors, low-contrast themes)

Step 1: Object Taxonomy

List every UI element the detector must classify, at minimum:

• Left bubble (incoming), right bubble (outgoing)
• Attached image / sticker / voice message
• Profile icon, sender name
• Timestamp, read-receipt marker
• System messages (e.g., "user joined", date dividers)

Step 2: OCR / Text Layer

• Language-specific OCR engine (e.g., Thai, English, mixed)
• Output: {bubble_id, bbox, text_string, sender_side, timestamp}
• Confidence threshold below which a bubble is flagged for human review rather than auto-cropped

Step 3: Anchor Keyword Dictionary

Define domain-specific start/end anchor terms. Example (financial evidence):

• Start anchors: account number, price agreement, "how much total"
• End anchors: "transferred", slip/receipt image, "payment sent"

For other domains (harassment, threats, contract agreement), swap the dictionary — the architecture stays the same, only the keyword set changes.

Step 4: Decision Flow (5 steps)

1. Object Detection — scan full long screenshot, produce bbox map for every bubble (x_min, y_min, x_max, y_max)
2. Text Content Analysis — OCR every detected bubble into text strings
3. Anchor Identification — run keyword decision tree to find Start Anchor bubble and End Anchor bubble
4. Auto-Drag Simulation — compute crop region:
   • y_top = y_start_anchor - 10px (margin)
   • y_bottom = y_end_anchor + 10px (margin)
   • Add ±1–2 bubble contextual buffer before/after the anchor range
5. Precision Crop Output — cut the original long image along the computed vertical band; discard unrelated content outside the band

Step 5: System Conditions

• Boundary Contrast Condition: bubble border/background must have sufficient contrast for reliable bbox detection; flag low-contrast themes for manual fallback
• Evidence Keyword Condition: maintain an extensible, domain-configurable keyword dictionary; anchor detection must fail safely (ask for human confirmation) if no clear anchor found
• Aspect Ratio Constraint: cropped output resized onto standard canvas (e.g., 645×890px) preserving original text proportions — no single-axis stretch/distortion

Step 6: Strict Prohibitions

• No bisecting bubbles: never cut through the middle of a bubble inside the selected range; if a bubble overflows the canvas, resize the canvas or push the whole bubble to the next page
• No context-free floating bubbles: never output an isolated bubble (e.g., only "transferred") without the preceding conversational context or referenced attachment — context-less evidence is legally weak
• No metadata stripping: never remove timestamps, read-receipts, or profile icons during crop — required for Chain-of-Custody integrity