LDC AI Prompt Playbook — Cross-Department Expectation Coverage Prompt Playbook ของ LDC — ครอบคลุมความคาดหวังทุกแผนก

Validate the attached file ({FILENAME}) before any analysis. Check and report: 1. Required columns: {LIST_REQUIRED_COLUMNS} 2. Missing values, especially in key fields ({KEY_FIELDS}) 3. Duplicate primary keys ({PK_FIELD}) 4. Date format consistency 5. Numeric outliers (Z-score > 3) per column 6. Broken formulas (#REF!, #VALUE!, #DIV/0!, #N/A) 7. Cross-tab inconsistencies (totals vs detail rows) Return a Green/Amber/Red findings table with: - Row count, missing count, duplicate count, outlier count - 3 example rows for each issue type - Recommendation: ready to analyze, fix needed, or send back to source Do NOT proceed to analysis until I confirm the data is clean.

Audit all formulas in the workbook ({FILENAME}). For each sheet, identify and report: 1. Formula errors (#REF!, #VALUE!, #NAME?, #DIV/0!, #N/A) with cell address 2. Broken sheet/file references (e.g., #REF! due to deleted tab) 3. Inconsistent SUMIFS/VLOOKUP ranges (e.g., one row uses A1:A100, next uses A1:A50) 4. Hardcoded numbers in formula cells (red flag for manual override) 5. Circular references 6. Formulas pointing to empty or hidden columns 7. Different formula logic in same column (data validation breach) Return a CSV-style audit log with columns: Sheet | Cell | Formula | Issue Type | Severity (High/Med/Low) | Suggested Fix Do NOT auto-fix anything. Only report. The Excel owner will approve fixes manually.

Based on the analysis above, create an action plan. For each finding (Red/Amber/Green), produce: 1. Problem statement (1 sentence, factual) 2. Root cause hypothesis (mark as "hypothesis" — not confirmed) 3. Three solution options 4. Recommended option with reason 5. Concrete steps (each with owner + due date placeholder) 6. Expected impact (qualitative or quantitative) 7. Review checkpoint (when and how to verify) Format as a markdown table: | Priority | Issue | Owner | Due Date | Action | Expected Impact | Review Date | Mark all owners as {OWNER_PLACEHOLDER} and dates as {DATE_PLACEHOLDER} for human assignment. Sort by Priority: Red first, then Amber, then Green.

Summarize this meeting from the transcript/notes below. Output structure: **Meeting:** {date, attendees, purpose} **Decisions:** (bullet list, factual only) **Action Items:** (table: Owner | Action | Due Date | Status) **Open Questions:** (questions raised but not resolved) **Risks/Blockers:** (anything flagged as risky or blocked) **Next Meeting:** (when, agenda, prep needed) **One-line Executive Summary:** (1 sentence for sharing in WhatsApp/Email) Rules: - Quote exact phrasing for decisions and commitments. - Do NOT invent action items not mentioned. - Mark unclear items as "needs clarification" instead of guessing. - Preserve technical terms (SBC, OOS, VDP, GP, MD, WO, SAR, etc.) exactly. [Paste transcript/notes here]

Use the attached Adjust_Stock and Lost_Over files. 1. Standardize Thai column names into a documented English schema. 2. Rank top 10 branches by negative variance amount (THB). 3. For top 3 branches, drill down by item, vendor, division, and % of branch sales. 4. Flag void/refund or PO records where the Z-score is >= 3 versus the same branch's 90-day baseline. 5. Generate a Red/Amber/Green alert queue with a severity score 0-100 (weight: amount 40%, frequency 30%, recency 20%, segment risk 10%). 6. For each Red alert, draft an investigation remark that lists what to check, which document to pull, and which staff role to interview. Mark every remark as "requires auditor confirmation". Output: - Branch ranking table (top 10). - Drill-down table for top 3 branches. - Alert queue table sorted by severity DESC. - One investigation remark paragraph per Red alert. Do NOT call any pattern "fraud". Do NOT name any individual employee in conclusions.

Using today's DC_Daily_Order, DC_Staff_Master, DC_Daily_Attendance, and DC_Stock_OnHand files: 1. Calculate priority rank per store: out-of-province first, then out-of-district, then by PCS volume DESC. 2. Compute required pickers per zone using: required_pickers = TotalLines / (BufferedHours * 30 lines/hr/picker) where BufferedHours = available shift hours minus a 15% buffer. 3. Filter eligible pickers — keep only attendance status Present, Late, or HalfDay. Exclude Absent and Leave. 4. Assign orders to pickers by zone match and capacity, balancing total lines per picker within +/-10%. 5. Flag stock shortages BEFORE picking starts: any order line where OrderQty > OnHand. Mark the affected store and item. 6. Output a supervisor action plan with: - Priority list (store, region, PCS, lines, due time). - Picker assignment table (picker, zone, assigned lines, capacity %, store list). - Shortage escalation list (store, item, OrderQty, OnHand, shortage qty, suggested action). - 3 escalation talking points for the morning huddle. Do NOT auto-send assignments to staff. The supervisor approves and dispatches.

Generate today's BA report pack from these files: Sales_All, Sales_Margin_ByItem, MarkdownWriteOff, Stock_OnHand_DOH, Targets_Period. Produce: 1. Daily sales by store with %Achievement, Diff_Target, and Avg_Per_Day (period-to-date). 2. GP analysis with GP (THB) and GP% per store and per category. 3. DOH per store and per item, highlighting the healthy range 25-35 days and flagging anything outside it. 4. MD + WO combined report with %Achievement vs the MD+WO target. 5. Green / Amber / Red flag per store using these rules: - GREEN: %Achievement >= 100% AND Growth (YoY or vs target) >= +5% AND GP% within +/- 2pp of plan. - AMBER: not Green and not Red. - RED: %Achievement < 80% for 2+ consecutive periods OR Growth <= -10% OR DOH > 60 days. 6. For each Red store, draft a one-paragraph root-cause hypothesis using only the data shown (e.g., shortfall driven by which category, which week, which item segment). Mark every hypothesis as "AI hypothesis — needs BA validation". Output as separate clearly-labeled tables and one flag summary at the top. Do NOT invent any number. If a metric is missing in source files, write "n/a — source missing" instead of estimating.