It's Tax Time: Building a Tax Document Assistant with Claude Code and the Ragie Skill

#ai #rag #claudecode #tutorial

It's tax time, which means a folder full of W-2s, 1099s, and IRS publications I'd rather not grep through manually. So I whipped up a tax doc assistant with our new Ragie skill — a CLI that answers questions from my actual tax documents with cited sources.

The stack:

Claude Code for the coding
The Ragie skill for RAG patterns (install with npx skills add ragieai/skills)
Ragie for document ingestion, retrieval, and metadata filtering
Claude Sonnet 4.6 for answer generation

What the Ragie skill gives you: context on the SDK. Ingestion patterns, retrieval methods, metadata filtering, rerank config. When I asked Claude Code to ""write an ingestion script that tags PDFs with type and year,"" it produced correct code on the first try because the skill had already loaded the relevant SDK knowledge.

The key patterns

Metadata tagging at ingestion. Each document gets classified at upload:

W-2s → type: "w2"
1099s → type: "1099"
IRS pubs → type: "irs-publication"
Receipts → type: "receipt"
Plus taxYear: 2025 on all of them

Type-scoped retrieval. Questions about freelance income only search 1099s. Questions about deduction eligibility only search IRS publications. The metadata filter narrows the search space before hybrid search and rerank run, which is a big quality win over ""search everything.""

Rerank on. Small latency cost, noticeable quality improvement.

Citations in generation. Every answer references the source document name, so you can verify anything the assistant tells you.