Description

Adds two Copilot skills under .github/skills/ in the azure-ai-contentunderstanding package that walk SDK users (and AI coding agents) through creating a custom analyzer end-to-end using the typed ContentUnderstandingClient already shipped in the package.

Zero public SDK API changes. Skills + scripts live under .github/, which is excluded from the PyPI sdist via the existing include-only MANIFEST.in — pip install azure-ai-contentunderstanding is byte-identical for consumers.

Each skill walks: env check → layout extraction → schema authoring (starting from a template) → local validation → analyzer create → batch test → category-aware stdout summary with leaf-level field rollout → optional ephemeral cleanup.

What's in this PR

End-to-end smoke runs

Both skills were executed against samples/sample_files/mixed_financial_docs.pdf (a packet containing invoice + bank statement + loan application). Ephemeral cleanup of all created analyzers confirmed for both runs.

Single analyzer (cu-sdk-generate-analyzer)

========================================================================
[SUMMARY]

category: (single)  (1 document)
--------------------------------
  field                                    fill rate   avg conf
  billToAddress                            100.0%      0.922
  customerId                               100.0%      0.953
  customerName                             100.0%      0.903
  dueDate                                  100.0%      0.979
  invoiceDate                              100.0%      0.977
  invoiceNumber                            100.0%      0.956
  lineItems[].amount                       100.0%      0.928
  lineItems[].date                         100.0%      0.902
  lineItems[].description                  100.0%      0.661
  lineItems[].itemCode                     100.0%      0.819
  lineItems[].quantity                     100.0%      0.932
  lineItems[].tax                          100.0%      0.924
  lineItems[].unitPrice                    100.0%      0.931
  poNumber                                 100.0%      0.961
  salesTax                                 100.0%      0.939
  shipToAddress                            100.0%      0.929
  subtotal                                 100.0%      0.917
  totalAmount                              100.0%      0.939
========================================================================

18 leaf rows reported, 100% fill rate across all extracted fields. Lowest confidence 0.661 on lineItems[].description — exactly the "agent reads the summary, proposes targeted v2 edits" loop the skill is designed for.

Classify-and-route (cu-sdk-generate-analyzer-classify-route)

========================================================================
[SUMMARY] (category-aware)

category: bank_statement  (1 segments)
  accountHolder                  100.0%      0.958
  accountNumber                  100.0%      0.962
  statementPeriod                100.0%      0.903

category: invoice  (1 segments)
  invoiceDate                    100.0%      0.979
  invoiceNumber                  100.0%      0.956
  totalAmount                    100.0%      0.872

category: loan_application  (1 segments)
  applicantDateOfBirth           100.0%      0.902
  loanAmountRequested            100.0%      0.886
  signatureDate                  100.0%      0.956
========================================================================

3 categories correctly identified, 9 fields × 100% fill. Per-category denominator verified (each category's fill rate counts only segments classified into that category, not packet-wide total).

All SDK Contribution checklist

• The pull request does not introduce [breaking changes] — zero public API changes; .github/ is not in the sdist
• CHANGELOG is updated for new features, bug fixes or other significant changes. — N/A: no shipped code changed; skills are dev-experience artefacts under .github/ and do not appear in any released wheel. Happy to add a "Other Changes" CHANGELOG entry if reviewers prefer.
• I have read the contribution guidelines.

General Guidelines and Best Practices

• Title of the pull request is clear and informative.
• There are a small number of commits, each of which have an informative message.

Testing Guidelines

• Pull request includes test coverage for the included changes. — 19 unit tests in tests/test_skills_*.py (validator purity, classifier wiring, prebuilt-routing passthrough, single-analyzer argparse + invalid-schema-pre-import + leaf-row summary). Helper scripts are thin wrappers over the typed SDK; the underlying API calls are already covered by the SDK's recorded tests, so the manual smoke run captured above covers the end-to-end path.

Verification steps

cd sdk/contentunderstanding/azure-ai-contentunderstanding

# Run the new unit tests (no Azure auth needed)
python -m pytest tests/test_skills_*.py -v
# expected: 19 passed

# Reproduce the single-analyzer smoke run (requires CONTENTUNDERSTANDING_ENDPOINT)
cp .github/skills/cu-sdk-generate-analyzer/templates/schema_template.json schemas/invoice_v1.json
# edit schemas/invoice_v1.json — replace REPLACE: placeholders
python .github/skills/cu-sdk-generate-analyzer/scripts/create_and_test.py \
    --schema schemas/invoice_v1.json \
    --input samples/sample_files/mixed_financial_docs.pdf \
    --output test_results/v1 --ephemeral