Agent Skills Highlights Summary
Every highlight bullet from the 10 pages in this section, gathered on one page and grouped by the page it came from.
- An Agent Skill is an operational contract for an AI assistant, not a tutorial for a human
- Triggers, an input/output contract, and guardrails are what make a skill invocable rather than just informative
- A skill narrows a model's default behavior to one team's Python conventions - it doesn't replace the model's underlying knowledge
- A skill's output still needs a verifiable, human-checkable command - pytest, ruff, mypy - not just plausible-looking code
- Scaffolds routers, Pydantic models, lifespan, and health routes
- Pins FastAPI 0.115+ and Pydantic 2 patterns
- Verification ends with ruff, mypy/pyright, pytest, and curl /health
- Never commits secrets - .env.example only
- Human cookbooks in fastapi section stay authoritative
- Defines extract, validate, transform, load stages with idempotent loads
- Pins pandas 2.2+ or Polars 1.x per team standard
- Outputs data contracts, Great Expectations/pandera stubs, and runbook
- Partitioned parquet and explicit schema evolution policy
- Verification includes row counts, null rates, and sample hashes
- Scaffolds train/eval loops with PyTorch 2.6+ and device autodetect
- Requires reproducibility seeds and version-pinned torch/cuda
- Outputs checkpointing, metric logging, and config YAML
- Guardrails against data leakage and train/eval mode bugs
- Verification includes loss decrease on synthetic batch
- Scaffolds ingest, chunk, embed, retrieve, and generate pipeline
- Requires evaluation set and citation grounding policy
- Outputs guardrails for PII, prompt injection, and tool auth
- Pins vector store and embedding model in config
- Verification includes golden questions and refusal cases
- Scaffolds src-layout package with pyproject.toml and uv
- Outputs README, LICENSE, CHANGELOG, and CI publish workflow
- Uses trusted publishing or API token via secrets - never in repo
- Verification: uv build, twine check, pytest, ruff
- Semver and tag-triggered releases documented
- Adds pytest layout, fixtures, markers, and parametrization
- Targets meaningful coverage on pure functions first
- Outputs conftest.py patterns and CI pytest command
- Discourages mocking httpx without contract tests
- Verification: pytest -q --cov with threshold gate
- Adds annotations incrementally with mypy/pyright clean modules
- Uses Python 3.14 syntax: union |, generics, Protocol
- Outputs pyproject tool.mypy or pyrightconfig.json
- Strangler pattern on legacy modules - not big-bang
- Verification: mypy --strict on touched package
- Scaffolds idempotent infra scripts with boto3 and typed wrappers
- Requires dry-run and change preview before apply
- Outputs IAM least-privilege policy stubs and audit logging
- Never stores long-lived access keys in repo
- Verification includes moto tests or LocalStack smoke
- Runs ruff format/check and applies PEP 8 idioms
- Flags mutable defaults, bare except, and pathlib opportunities
- Outputs prioritized refactor list with risk tiers
- Never changes behavior without tests
- Verification: ruff, pytest, optional mypy on touched modules
Revisado por Chris St. John·Última actualización: 31 jul 2026