When building AI agents that interact with real-world systems—filling tax forms, checking blockchain RPCs, or normalizing sensor units—prompting an LLM to "figure it out" at runtime is fragile.
Skillware (github.com/ARPAHLS/skillware, MIT) takes a different approach: package each capability as a self-contained, versioned Python bundle with a typed contract (manifest.yaml), deterministic code (skill.py), agent instructions (instructions.md), and offline unit tests (test_skill.py).
Today we released v0.6.2 (pip install skillware==0.6.2). Here's a quick look at what's inside:
1. Filling Tax & FATCA PDFs Without Context Bloat (office/pdf_form_filler v0.2.0)
Official IRS PDFs like W-8BEN-E have 179+ widgets and often ship with hybrid XFA wrappers that break standard PDF libraries.
In v0.6.2, office/pdf_form_filler adds:
- 4 verified static AcroForm templates (
us_irs_w8ben,us_irs_w9,us_irs_w8bene,us_irs_f8966_fatca) with XFA stripped. - Section-chunked YAML schemas: Call
get_template_schema(template_id="us_irs_w8ben", section="part_1_identification")so the model only sees one section at a time. - Automatic PII masking: Fields tagged
sensitive: true(TINs, birthdates, account numbers) are masked (***4321) inconfirmed=Falsepreviews. - 3-tier operator overrides: Override bundled templates via
PDF_TEMPLATES_DIR,.skillware.yaml, or~/.config/skillware/pdf_templates/.
Run the offline demo yourself:
pip install "skillware[office_pdf_form_filler]" python examples/pdf_form_filler_w8ben_demo.py 2. Pre-Flight EVM RPC Health Gate (monitoring/rpc_health_probe)
Before an agent signs or broadcasts an on-chain transaction, monitoring/rpc_health_probe checks eth_chainId, round-trip latency, and optional reference block lag—failing closed on wrong chains or JSON-RPC errors.
3. Deterministic Unit Normalization (data_engineering/unit_normalizer)
Dependency-free conversions for single values, batches, and tables—including compound strings (5 ft 11 in), temperature offsets, and explicit ambiguity flags (ton, gal, oz).
Get Involved
Skillware is built for both human contributors and AI-assisted workflows (docs/contributing/ai_native_workflow.md). Check out our good first issue list on GitHub!



