Claude Code invoice skill splits AI from the math

The author frames the whole exercise around three concrete failures that show up when you ask Claude Code for a business document. Line items multiply and the subtotal or consumption tax can drift, and a wrong invoice damages the relationship with the client. The layout shifts every time, with different margins and typefaces from the previous version. And because the AI cannot carry a number across separate conversations, document numbers collide.

The published setup answers each of those by moving responsibility rather than improving the prompt. Drafts are rendered with a 'draft' mark in the top left and no number at all, and the script counts whether content overflows the A4 page in Chrome's PDF output and stops if it does, pushing overflow to the next page automatically when there are many line items. Issued documents are recorded line by line in a file, the next number is taken after the largest one in the same month, and issued documents are never overwritten.

The author's general rule, drawn from building this, is that anything where a mistake is costly and anything that should be identical every time belongs in a script, while the AI is left only to read the intent out of the conversation. The stated reason is that instructions in a SKILL.md file — round down on consumption tax, number sequentially — are obeyed only by luck, whereas a script with tests behaves the same way each run.