01
Check
The Claude skills or plugin you already use, checked.
- Validation and a linter for silent failures
- A routing test on the phrases your team uses
- A short report: what misfires, and why
Claude skills
Claude skills for your month-end numbers, tested on your own export.
Claude is good at arithmetic; what goes wrong is what it counts: test orders, VAT, which month a refund or a payout belongs to. I build Claude skills that compute every figure with a script, name every order left out and print your definitions under every answer. Or I check the skills or plugin you already use.
What was tested
Three synthetic Shopify stores, nine owner questions, three runs each with Opus 5.5, with the skills and without. Every answer is public.
27/27answers on the stated basis with the skills; 13/27 for plain Claude
27/27named everything the owner should see; 21/27 without the skills
6/6answers named every excluded order on a 4,470-order store; 0/6 without
8/8export changes stopped or named; the earlier version let 6 through silently
What you can order
Fixed price, quoted per export.
02
One skill
One month-end skill built on your export.
- A sales summary, payouts to the bank or product margins
- Your definitions asked once, printed under every answer
- Tested on your recent months, three runs each
03
Month-end pack
All three skills, with every check.
- Summary, payouts and margins
- Every figure traced to the script, everything left out named
- Mistakes planted in your data; every answer and grade handed over
How it works
You send a question and an export
The question your team asks Claude every month, and a sample export. Test data or an anonymised copy is fine.
A first run on your export
You see what was counted and what was left out before anything is built. The skills have not yet run on a real store's raw export, so this step comes first.
Built and tested
Three runs per question, mistakes planted on purpose, every answer read.
You get the skills
A plugin for Claude Code, its tests, every answer and grade, and install notes.
Cases
Claude skills27/27answers on the stated basis with the skills; 13/27 for plain Claude
Claude month-end reports for Shopify that show their work
Read the case → Claude plugin check12/12planted problems found in every close; the revenue line changed from run to runAnthropic's free Small Business plugin, checked on synthetic books
Read the case →Synthetic stores, Opus 5.5 only, three runs per question: a repeated behaviour, not an error rate. The same person wrote the skills and the tests; every answer and grade is public, so anyone can re-grade them. Your files are read by Claude for your order only and deleted after delivery unless you ask me to keep them.