Claude Ecom

# Claude Ecom

Use this for an e-commerce store review, store health check, or questions over order data like "how was last month", "how is retention", or "why did revenue drop".

## Input
An order transaction CSV, one row per order or line item. It needs an order ID, order date, customer ID or email, and revenue after discounts. Quantity, SKU, and discount columns help. If no file is named, look for CSVs in the folder and ask if there is more than one candidate.

## Pick the mode
- No arguments or "review": a full review that picks periods from the data and writes REVIEW.md.
- "review 30d", "90d", or "365d": a full review for one period.
- A natural-language question: a focused answer of 10 to 30 lines, inline, no file.

## Phase 1: compute
Run the bundled engine on the CSV. It computes KPI trees, about 30 pass/watch/fail health checks, and the 30, 90, and 365-day windows, and saves the results as JSON. The first run sets up a private Python environment.

## Phase 2: interpret
Python computes the numbers and you explain them. Periods tell you where things are heading; health checks tell you what is broken now. Combine both into one story.

## Report sections, in this order
1. Executive summary: a 4 to 6 line narrative and a scoreboard table
2. 30-day pulse: KPI tree, at most 1 finding
3. 90-day momentum: KPI tree, drivers, at most 2 findings
4. 365-day structure: KPI tree, drivers, at most 3 findings
5. Action plan: at most 5 items by time horizon, ending with 2 or 3 guardrail metrics that must not get worse
6. Data notes: 2 to 4 lines

## Rules
- Every number comes from the engine output. Never quote outside sources.
- Each finding goes: what is, why it matters, what to do. Do not write "consider" or "optimize".
- Never repeat a finding across periods. Aim for about 150 lines.
- Flag data-quality warnings and do not treat a partial month as a real trend.

Adapted from @takechanman1228's approach: https://github.com/takechanman1228/claude-ecom