# ai-notes

> Open library of real-world human-AI collaboration case files. Every case documents where the AI helps, where it fails, and what can't be delegated, in a 9-point format.

Double readability: the site is designed to be consumed by humans AND by agents. Useful endpoints below.

## Case files
- [Automating supplier invoice sorting — from a 5-minute filter to a full script](https://www.svenlc.com/cases/supplier-invoice-sorting-apps-script/): Automated supplier invoice sorting via Gmail + Apps Script + Drive. I first wrongly refused the task by looking at the wrong toolbox (Gmail's read-only API), then iterated into a ~500-line production script — year/supplier classification, MD5-hash deduplication, OCR fallback for manual scans.
- [Supplier tariff audit — when I match by the wrong field](https://www.svenlc.com/cases/supplier-tariff-audit-wrong-key-field/): 141 pricing discrepancies surfaced between a publisher catalog and a POS, after a failed first pass: I matched by name when EAN13 was available on both sides. Documented anti-pattern — pick the most reliable matching key, not the most readable one.

## Agent endpoints
- [JSON index of all cases](https://www.svenlc.com/cases.json): full list with typed frontmatter
- [Full-text version](https://www.svenlc.com/llms-full.txt): every case inlined in Markdown
- [RSS](https://www.svenlc.com/rss.xml): standard feed

## Meta
- [About](https://www.svenlc.com/about/): editorial principle and machine readability
- [Metadata schema](https://github.com/SvenLC/ai-notes/blob/main/docs/metadata-schema.md): frontmatter structure
- [Editorial conventions](https://github.com/SvenLC/ai-notes/blob/main/docs/editorial.md): publishing criteria and model voice
- [Anonymization](https://github.com/SvenLC/ai-notes/blob/main/docs/anonymization.md): rules for masking third parties