Cases · Claude plugin check
Anthropic's free Small Business plugin, checked on synthetic books
Self-initiated case study on synthetic books and three synthetic Shopify stores, Opus 5.5, three runs per question. An independent check, not affiliated with or endorsed by Anthropic.
Goal
Many small businesses will install Anthropic's free plugin rather than pay for skills of their own. The question: where does it already do the job, and what does the owner still not see?
Challenge
The plugin's 44 skills are written as instructions, so the model writes its own code each time. Where an answer rests on a rule the owner never gave, the plugin picks one, and a fresh start can pick another; the arithmetic is right either way.
Solution
The plugin closed August on books with 12 planted problems, three runs; answered the nine owner questions of my bench with its closest skill called by name; and was routed on plain phrasings, a fresh session per message. Every answer was kept, read and compared with plain Claude and with my own skills on the same stores.
Outcome
It found all 12 problems in every run and kept its own receipt and duplicate rules, where plain Claude chased a 14.20 coffee receipt in all three runs. But the corrected P&L gave revenue of 96,459.17, 94,396.54 and 96,680.77 in three runs; one run counted an order tagged "test" that two left out; and asked "How did August go?" with the export in the folder, its skills ran in 0 of 27 sessions. None of this is a bug: it is the part a general plugin cannot know about one business.
Evidence
Every answer, workbook and reading (GitHub)
The plugin itself is Anthropic's, under Apache-2.0, and is not in the repository.
From the case
The case, 2 pages (PDF)
Service
Claude skills for your own files
Your counting rules asked once, checked on your exports, tried on the way your team asks.