Litestar styleguide
Skill litestar-org/litestar-skills/plugins/litestar/skills/litestar-styleguide
Use when authoring Litestar skill content, Python/TypeScript examples, PEP 604, async I/O, Google docstrings, ruff/mypy/pyright, pytest, or CI rules. Not for focused Litestar APIs.From its SKILL.md
npx -y skills add litestar-org/litestar-skills --skill litestar-styleguideAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 13 stars13 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
- runs commandsInstructs the agent to run 1 command, including `check via grep when editing`.
SKILL.md
5.2 KB, ~1.2k tokens by cl100k_base, as published. Nobody here has run it
litestar-styleguide
This is the shared style baseline that every other skill in this plugin references. It exists so that cross-cutting rules (PEP 604 unions, async I/O, ruff + mypy + pyright, test file naming, CI/CD conventions) live in exactly one place — and individual skills stay focused on their framework or tool-specific surface.
What's in here
Each reference covers one slice of the code-style baseline. A sibling skill links to only the files relevant to its language / framework mix.
references/general.md— Cross-language principles (simplicity over cleverness, error handling, naming, DRY-when-it-pays)references/python.md— Python conventions (PEP 604 unions, docstrings, imports, ruff / mypy / pyright configuration)references/litestar.md— Litestar-specific baseline (guards, DTOs, DI, plugin use)references/typescript.md— TypeScript conventions (when paired with a frontend skill)references/testing.md— Testing conventions (pytest, vitest, fixtures, coverage)references/ci-cd.md— CI/CD conventions (GitHub Actions, matrix builds, caching)
How sibling skills consume this
Every SKILL.md in this plugin has a ## Shared Styleguide Baseline section near the bottom. That section links to a subset of these references — only the ones that apply to the skill's language / framework mix. For example:
skills/litestar/SKILL.mdlinks togeneral.md+python.md+litestar.mdskills/litestar-vite/SKILL.mdlinks togeneral.md+typescript.md+litestar.mdskills/litestar-testing/SKILL.mdlinks togeneral.md+testing.md+python.md+litestar.md
The sibling skill extends the baseline with its own tool-specific Code Style Rules, Quick Reference, Guardrails, and Validation — but it does not duplicate the baseline. If a convention is generic (type hints, naming, imports), it belongs here.
When to update this skill
- A rule becomes contentious across two or more sibling skills → pull it into the right baseline reference file here.
- A new language lands (Rust, Mojo, etc.) → add a new
references/<lang>.mdand link from skills that use it. - A tool is swapped out (e.g., ruff replaces flake8 + black) → update
python.mdonce; all sibling skills inherit it.
Authoring rule for this skill
- Keep references terse, imperative, authoritative. No hedging ("you might want to…"). State the preferred choice and the one-line reason.
- Every "never do X" rule has a one-line why (perf, runtime introspection, OpenAPI alignment, etc.). No folklore.
- Examples are copy-pasteable and minimal. No pseudo-code.
Workflow — consuming this baseline
- Open the sibling skill you are editing (
skills/<name>/SKILL.md). - Look at its
## Shared Styleguide Baselinesection — it already lists a subset of the references here. - When adding a rule to the sibling, ask: is it generic (language/tooling) or framework-specific? Generic → land it in the right file under
references/here. Specific → keep it in the sibling. - Cross-link bidirectionally if a rule here is amplified in the sibling.
Guardrails
- No duplication across skills. A rule lives in exactly one file; sibling skills link to it.
- No folklore. Every rule has a one-line justification (perf, runtime introspection, OpenAPI alignment, etc.). Delete rules you cannot justify.
- Terse and imperative. Bullets are ≤ 2 sentences. If a topic needs more, split it into its own reference file.
- Examples are minimal and copy-pasteable. No pseudo-code; no multi-hundred-line fixtures.
Validation Checkpoint
- Every sibling skill's
## Shared Styleguide Baselinesection resolves to files that exist underreferences/ - No rule is duplicated between two reference files (check via grep when editing)
- Each "never do X" rule has a one-line
Reason:explanation - New language support lands as a single new
references/<lang>.md— not scattered into sibling skills
Example — adding a new rule
A reviewer finds that two sibling skills independently wrote "use ruff format not black". Instead of leaving duplicates, pull the rule into references/python.md:
- **Use `ruff format`, never `black`.** Reason: ruff is the single toolchain for
lint + format; running two formatters produces style drift.
Then in each sibling's SKILL.md, replace the duplicate with a pointer:
## Shared Styleguide Baseline
- [Python](../litestar-styleguide/references/python.md)
</example>
Official References
- https://peps.python.org/pep-0604/ — PEP 604 union syntax
- https://docs.astral.sh/ruff/ — ruff linter / formatter
- https://microsoft.github.io/pyright/ — pyright type checker
- https://docs.pytest.org/ — pytest
What ships with it: 8 files
42.0 KB alongside SKILL.md
references/
- canonical-apps.md2.4 KB
- ci-cd.md3.6 KB
- general.md2.8 KB
- google-developer-knowledge-mcp.md3.6 KB
- litestar.md10.3 KB
- python.md7.8 KB
- testing.md5.9 KB
- typescript.md5.5 KB