Python prompt.

pythonprompt.md

python coding tips for claude code etc

• Keep explanations brief and to the point.
• I am mostly familiar with Bash, Stata, Python, and Go.
• Do not provide code snippets and examples of code unless asked.

Looks

• Do not use emojis.

Frontend

• Prefer Cloudflare Pages and HTMX + Tailwind CSS.
• Avoid npm/webpack/JS builds.

Infrastructure

• Cloudflare (domains, R2, D1, Workers, Containers, Pages, AI Gateway, Queues, KV, Calls/RealtimeKit).
• Use D1 unless there are limitations or need Postgres-specific features (JSONB, FTS, PostGIS).
• While Cloudflare Workers support Python also consider whether Typescript workers are more reliable.
• When creating R2 buckets, enable Data Catalog explicitly (not default).
• Multiple cloud accounts/domains may exist across providers.
• For low latency, prefer WebRTC (for audio/video), WebSockets, SSE, UDP, over REST.

Python coding

• Use Python 3.12
• uv, uv workspaces, ty (LSP), pytest, ruff, type hints on all functions.
• FastAPI + SQLModel (Pydantic + SQLAlchemy combined) for APIs and databases.
• Prefer serverless approaches that work in Containers or Workers over virtual machines.
• Always try to use SQLModel and fallback to python dataclasses if necessary.
• Use exchange_calendars for trading holidays by exchange.
• Concise comments explaining intent. Tests with mocks/fixtures. 4-space indent.
• Be mindful of timezones; Modal.com uses UTC/GMT.

Python analytics

• Use gymnasium, not OpenAI gym (deprecated).
• Prefer polars over pandas; use numpy for numerical compute.
• Store tabular data as parquet (zstd compression if available).
• DuckDB for analytical SQL queries on parquet/polars. dbt for SQL.
• Prefer hardware-accelerated libraries: polars (SIMD), numpy (BLAS/LAPACK), cuDF/RAPIDS (GPU).
• Prefer Prefect for orchestration and workflows over Airflow.

Python performance

• Prefer async/await for I/O-bound work (HTTP, DB, file).
• Use multiprocessing or ProcessPoolExecutor for CPU-bound (GIL bypass).
• Profile first (cProfile, py-spy, scalene) — don't guess bottlenecks.
• Batch external calls (DB queries, API requests) — N+1 is the common killer.
• Connection pooling for databases and HTTP (httpx, asyncpg, sqlalchemy pool).
• Use generators/iterators for large datasets — avoid materializing full lists.
• Prefer list/dict comprehensions over loops — faster and clearer.
• NumPy/polars vectorization over Python loops for numerical work.
• functools.lru_cache or @cache for expensive pure functions.
• slots on data classes with many instances (reduces memory ~40%).
• Avoid repeated attribute lookups in tight loops — assign to local variable.
• f-strings are fastest for string formatting.
• orjson for JSON (10-20× faster, handles serialize + parse).
• pysimdjson only if parsing large (>1MB) JSON is the bottleneck.
• uvloop as event loop for async workloads (2-4× faster than default).
• Struct packing for binary protocols — struct module or msgspec.
• Memory views for zero-copy buffer manipulation.

LLMs

• Use massive.com (formerly Polygon.io). There is already a running R2 bucket ('warrenbucket') with historical data.
• GPU Service: Modal.com (primary) + RunPod.io (backup)
• Use huggingface.co
• Vector DB: Locally use Qdrant. Keep in mind potential sync in later iterations to Cloudflare Vectorize.
• Use a versioned, compatible embedding library that can be used in both local development and Cloudflare Vectorize.
• Use wandb for logging and debugging language model inputs, outputs, and traces.

Panel data and regression approaches

• Prefer linear probability model over logit; always use robust (HC) standard errors.
• Panel fixed effects: use linearmodels (PanelOLS) -- better than statsmodels for clustered SEs and IV.
• Mixed effects (statsmodels.MixedLM): hierarchical data with random intercepts/slopes (repeated measures, nested structures like students within schools). Use when group-level variance matters or predicting for new groups.
• Use fixed effects over mixed effects when within-group variation is what identifies causal effects.