Claude Ecom — Ecommerce Business Review
Turn an orders CSV into a consultant-grade business review in one command: KPI decomposition, ~30 health checks, prioritized findings, and next actions.
Tool permissions are least-privilege declared (Read/Grep/Glob/Write and scoped Bash(ecom *)); data flow is transparent: all numbers must come from review., external sources are forbidden, output is a deletable REVIEW.md. Deductions: first run bootstraps a private venv under ~/.local/share/claude-ecom (writes to the user home without explicit confirmation), the bundled Python engine and launcher source are not present in the evidence and cannot be statically audited; publisher identity is unverified, attribution limited.
Instructions are internally consistent: mode selection, report contract, data rules, check definitions and cluster activation rules align, with failure feedback specified (fallback launcher if ecom is not on PATH, degraded answers when 365d coverage is missing). Deductions: the Python engine itself is absent from the evidence so the happy path cannot be statically reproduced; fuzzy CSV column matching, abnormal-input error behavior and dependency availability have no tests or sample data.
Triggers are explicit (ecommerce review, store health, business review), the argument-hint and mode table are clear, input requirements (required CSV columns) are stated, and language rules require matching the user's language (good for Chinese users). Deductions: benchmarks are USD-based with Western ecommerce calendars, no localization for Chinese e-commerce scenarios; non-fit ranges (B2B, marketplace) are only briefly noted in benchmarks.md; trigger boundary declarations are incomplete.
Documentation is well layered: SKILL.md entry point plus on-demand references with clear progressive disclosure, data rules, quality gates and prohibitions. Deductions: MIT license is repository-level metadata not restated in the skill; no version number, changelog, or maintainer/update path; some referenced files (e.g. review-schema.md) were not included in the evidence; dependency/install notes rely on an unseen bin/ecom launcher.
The target scenario is clear and the design is mature: the two-phase split (Python computes, Claude interprets) is sound, the report contract enforces structured, directly usable output, and data rules reduce hallucinated numbers. Deductions: static review cannot verify actual review. quality or benchmark correctness; core value depends on the unseen Python engine; no representative real output is verified, only templates and worked examples.
References cite third-party sources (Bain, Shopify, Recharge, BS&Co) and provide auditable formulas and threshold definitions, with fact/inference separation (confidence labels like Medium-Low). Deductions: no tests, no CI evidence, no executable reproduction material; several benchmark sources are blog/marketing statistics of mixed reliability; some referenced files are missing from the evidence, so not all claims can be cross-checked.
- The bundled Python engine and bin/ecom launcher source are not in the reviewed evidence; audit them before use, especially the first-run venv bootstrap under ~/.local/share/claude-ecom/.
- Benchmarks are USD-denominated and based on Western ecommerce calendars; recalibrate thresholds for Chinese/local e-commerce scenarios.
- No tests or CI evidence; KPI computation correctness is not independently reproduced — spot-check key conclusions manually.
- Publisher identity is unverified by FollowSkills; license and maintenance path are repository-level metadata only, long-term maintenance status unknown.
What does this skill do, and when should you use it?
This Claude Code skill targets D2C ecommerce stores. It reads an order-transaction CSV, runs a bundled Python engine that computes KPIs, new-vs-returning KPI trees, revenue driver decomposition, and ~30 pass/watch/fail health checks across 30d/90d/365d windows. Python owns the numbers; Claude interprets them and writes a REVIEW.md with business context, or answers a focused question inline. MIT-licensed, and the core engine also works standalone via pip (producing review. and a basic report without the narrative layer).
Reads an orders CSV (required columns: order ID, order date, customer ID or email, revenue after discounts; optional: quantity, SKU, discount amount — column names are fuzzy-matched). Runs the ecom CLI (first run bootstraps a private Python venv under ~/.local/share/claude-ecom/) to produce review. containing: per-period summary KPIs with prior-period comparison, new-vs-returning KPI trees, revenue driver decomposition (AOV/volume/mix), 365d repeat purchase rate and 12-month trend, and ~30 health checks scored 🟢/🟡/🔴. Claude then writes REVIEW.md in a fixed structure: Executive Summary, per-period Pulse/Momentum/What-is→Why-it-matters→What-to-do sections, an Action Plan with deadlines, success metrics and Guardrails, and Data Notes.
- A D2C brand owner without a staff analyst who writes monthly business reviews from scratch and wants a one-command report draft
- An ecommerce manager who knows revenue dropped but can't explain why, needing KPI trees and driver decomposition to isolate AOV vs volume vs mix
- A data analyst or marketer automating monthly store health checks, delegating repetitive computation to Python
- A store owner who wants quick answers like 'how was last month', 'how's retention', or 'why did revenue drop' as short inline replies instead of full reports
- A retail manager wanting a consistent 🟢/🟡/🔴 signal system across Revenue, Customer, and Product dimensions
What are this skill's strengths and limitations?
- Clean separation of concerns: Python computes numbers for precision, Claude writes the narrative for readability
- Strictly structured output: KPI trees, 30/90/365d multi-horizon view, 'What is → Why it matters → What to do' findings, action plan with guardrails
- Explicit data rules: all numbers must come from review., external sources are banned, preventing hallucinated figures
- Core engine is independently usable via pip (claude-ecom) — works without Claude Code for basic reports
- Validated on the UCI Online Retail II dataset (~1M real transactions) with a full example report published
- Clear principles for incomplete data: omit what can't be measured, no N/A or empty placeholder sections
- Input is orders-transaction CSV only; Shopify API integration is on the roadmap but not implemented
- Report format is rigid (fixed section order, line and finding-count caps) — users wanting free-form output may feel constrained
- Health checks use generic D2C thresholds that may not fit specific categories or business models
- Publicly validated on only one dataset (Online Retail II); no test suite or cross-platform evidence provided
- First run bootstraps a venv and may take a minute; the focused-query period mapping depends on reference docs with unclear customization room
- Relies on Claude Code-specific mechanisms ($ARGUMENTS injection, allowed-tools Bash permission syntax), so porting to other Agent Skills clients requires edits
How do you install this skill?
Prerequisites: Claude Code and Python 3.10+. Install as a Claude Code plugin: run /plugin marketplace add takechanman1228/claude-ecom, then /plugin install claude-ecom@claude-ecom. The Python backend auto-installs into a private venv under ~/.local/share/claude-ecom/ on session start. If migrating from v0.1.x (curl installer), remove the old copy with rm -rf ~/.claude/skills/ecom to avoid duplicate skills. The skill file lives at skills/ecom/SKILL.md in the repo.
How do you use this skill?
Drop your orders CSV into the project (if unspecified, the skill Globs for *.csv and asks which to use when multiple candidates exist). Then run: /claude-ecom:ecom review (full review, auto-selects periods, outputs REVIEW.md); /claude-ecom:ecom review 30d / 90d / 365d (single period, outputs REVIEW_{PERIOD}.md); or /claude-ecom:ecom review How's retention? (focused question, inline answer, no file). Plain-language prompts like 'review my store' also trigger it automatically. CLI-only usage (no Claude Code): pip install claude-ecom, then ecom review orders.csv.
How does this skill compare with similar options?
The repo explicitly credits claude-ads by @AgriciDaniel as inspiration — that skill focuses on ad analysis, while claude-ecom targets order/sales data for business reviews. If you want only the compute layer without the narrative, use the standalone claude-ecom package on PyPI.