![\"Comic](\"https:\/\/engel-wolf.com\/wp-content\/uploads\/tech-ynab-amazon-categorizer-01-bookkeeping-mystery-transactions.jpg\")
The problem: no API anywhere<\/h3>\n\n\n\n
Every month I had between 10 and 30 of these mystery transactions. For each one I opened Amazon or PayPal, searched the order, looked at the items and went back to YNAB. Ten to twenty minutes in total, and it felt pointless for a small charge.<\/p>\n\n\n\n
The obvious fix would be an API. Amazon: no API for your own orders. PayPal: nothing useful for this. The small online shop: forget it. But every one of them sends an order confirmation e-mail<\/strong>. That was the best idea in the whole project: forget the APIs, use e-mail<\/strong>.<\/p>\n\n\n\n Everything runs on a Raspberry Pi in Docker. There are two pieces:<\/p>\n\n\n\n The importer decides what counts as Amazon with a simple rule: the payee matches This is the heart of the tool, from A few things in there I would recommend to anyone. Give the model your real category lis<\/strong>t instead of letting it invent categories. Put IDs next to the names,<\/strong> so the answer is machine-usable without fuzzy matching. And spell out the ugly parts of the input: Amazon product titles start with seller codes and ASIN noise, and „remove cryptic codes“<\/strong> plus a hard length limit is what turns a 200-character title into something that fits in a YNAB memo. The service returns plain JSON: order number, category ID, category name, up to three products. The memo becomes A tool that calls an LLM has a dependency that expires. The categorizer broke in three ways, and each one failed silently<\/strong>:<\/p>\n\n\n\n Lesson: give the agent the list of suspects, and write down what you learned.<\/p>\n\n\n\n The biggest takeaway is not the fix, it is the runbook. Claude wrote an Because a silent failure was the real problem, the same session added a small monitoring card to the dashboard. The Pi has under a gigabyte of RAM, so Grafana was out. Instead: a tiny standard-library status service that scans container logs for errors, plus Dozzle<\/strong> as a log viewer, both as ready-made ARM images, because building Python containers on that Pi is pain. When I asked whether Dozzle was overkill, Claude argued for keeping it for context around an error, and when I claimed an environment variable would filter log lines, it checked Dozzle’s own help output instead of agreeing with me. It was right; the setting I actually wanted was the log level, now set to WARN.<\/p>\n\n\n\n A monthly chore that used to eat 10 to 20 minutes, sometimes more, now takes about two. Not a spectacular demo, just a small agent that removes a repetitive annoyance from everyday life. And the second lesson, maybe the more important one: a small agent needs a small amount of care, and the best care is a runbook the next agent can read.<\/p>\n\n","protected":false},"excerpt":{"rendered":" Order e-mails, IMAP, one Claude prompt and YNAB: how a ten-minute side project turned cryptic Amazon payments into readable, categorized transactions, why it silently broke when the model was retired, and why AGENTS.md is now the most important file on my Raspberry Pi. Weiterlesen The setup<\/h3>\n\n\n\n
POST \/categorize<\/code>. Order e-mails are forwarded to a dedicated mailbox, and the service does 3 steps: <\/li>amazon|amzn<\/code>, or the memo contains something shaped like an Amazon order number (\\d{3}-\\d{7}-\\d{7}<\/code>). For those it calls the categorizer and rewrites the memo. Instead of a cryptic reference I now see something like „Shimano rear light \u2192 Cycling“ or „10-pack dishcloths \u2192 Household“, and the category is already proposed. I only approve or re-categorize.<\/p>\n\n\n\nThe prompt that does the work<\/h3>\n\n\n\n
claude_service.py<\/code>. The category tree from YNAB goes in, grouped and with IDs, so Claude can answer with an ID I can write straight back:<\/p>\n\n\n\n# Format tree for Claude (Group > Category with IDs)\ncategory_list = ""\nfor group, cats in sorted(category_tree.items()):\n category_list += f"\\n{group}:\\n"\n for cat in sorted(cats, key=lambda x: x['name']):\n category_list += f" - {cat['name']} [ID: {cat['id']}]\\n"\n\nuser_message = f"""Analyze this Amazon order email and suggest the best YNAB category.\n\nOrder Number: {order_number}\n\nEMAIL CONTENT:\n{email_body}\n\nYOUR YNAB CATEGORIES (organized by group):\n{category_list}\n\nPRODUCT EXTRACTION RULES:\n- Remove cryptic codes at the start (like SFGSUP, ASINs, SKUs)\n- Keep the meaningful product description\n- Truncate to max 10 characters per product name\n- Extract up to 3 main products\n\nEXAMPLES:\n[\u2026]"""\n<\/pre><\/div>\n\n\nAmazon Order <nr>: <products> - <category><\/code>.<\/p>\n\n\n\n![\"Comic](\"https:\/\/engel-wolf.com\/wp-content\/uploads\/tech-ynab-amazon-categorizer-02-bookkeeping-robot.jpg\")
What went wrong: the model ages, the code doesn’t<\/h3>\n\n\n\n
AGENTS.md: the most useful file in the project<\/h3>\n\n\n\n
AGENTS.md<\/code> on the Pi with „Rule #1: never rebuild these images“ at the top, the container layout, the import flow, the deploy-by-overlay recipe, a debugging checklist and a dated history of what broke. Short pointer files in each sub-project link back to it, and a memory note on my laptop says „read the runbook first“. The next agent that touches this stack, or me in six months, starts with the knowledge instead of rediscovering it. One line from that checklist I now apply everywhere: check what is in the image<\/em>, not only what is on disk.<\/p>\n\n\n\n![\"Comic](\"https:\/\/engel-wolf.com\/wp-content\/uploads\/tech-ynab-amazon-categorizer-03-bookkeeping-done.jpg\")
If you want to build this<\/h3>\n\n\n\n