尊厳あるPython：LLMエージェントを改善する10のルール

Meet Compass — Dagster’s new AI data analyst for Slack. Turn questions into trusted insights, instantly. Try Compass now →Discover What assets do best, an animated, narrated story about how data assets work together. Watch now →

Try Dagster+Sign InModern LLMs generate patterns, not principles. Dignified Python gives agents the intent they lack, ensuring code is explicit, consistent, and engineered with care. Here are ten rules from our Claude prompt.

Modern LLMs are trained on an enormous and unruly mix of code: questionable StackOverflow snippets, half-finished projects, and hobbyist repositories. When agents generate code from that background, the results can be fast but unfocused. Even when you don’t fully defer to them, the constant cycle of nudging, correcting, and rewriting can make development feel fragmented rather than cohesive.

At Dagster, intent has always mattered. The platform was built so engineers never have to guess what a system is doing or why it behaves the way it does. As we’ve adopted agent-assisted development, we’ve seen the same need emerge in code generation. LLMs are excellent at reproducing patterns, but they rarely understand the principles that make a codebase feel deliberate and unified.

Python’s flexibility amplifies this problem. There are often many valid ways to solve the same problem, and without guidance an LLM will default to whichever pattern it has seen most often. The result is code that technically works, but feels inconsistent. And can lead to a codebase that feels like it is authors with very different philosophies.

To address this, we codified our opinions about how Python should be written into a set of rules we call Dignified Python. Rather than relying on post-hoc cleanup through reviews or rewrites, we load these rules directly into the model’s context from the start. This gives agents a clear sense of our standards, conventions, and design philosophy.

By encoding intent upfront, we get the speed benefits of LLMs without sacrificing the clarity and consistency that make a shared codebase feel like a single system.

At Dagster we have strong opinions about how Python should be written. Historically we enforced these through linting and code reviews, but as agents have become more ingrained in our workflows we have codified a set of rules that we can supply directly as context when working with an agent. We refer to these rules as "Dignified Python."

Python's flexibility is one of its greatest strengths, but that flexibility becomes a liability when agents are involved. There are often many valid ways to accomplish the same task, and without guidance an LLM will pick whichever pattern it has seen most frequently. The result is inconsistent code that feels like it was written by a dozen different people with a dozen different philosophies.

Being opinionated cuts through this. By defining clear guidelines and loading them into Claude's context, we have drastically accelerated our engineering velocity while maintaining the quality and consistency of our codebase. The key insight is that you must impart your expertise to the model. An LLM has no understanding of your team's conventions, your architectural decisions, or the hard-won lessons that shaped your codebase. By encoding that expertise into rules, you ensure the AI produces code that reflects your standards rather than generic patterns.

The goal of Dignified Python is not just to generate code that works, it is to generate code that communicates intent. That is where LLMs often fall short. Models are excellent at reproducing patterns, but far less adept at expressing the design principles behind them. Intent is not about choosing the right snippet; it is about applying a coherent philosophy for how Python should be used. Our rules ensure that the code produced by agents reflects that philosophy.

Ten Rules from Our Claude Prompt

Below are ten rules extracted from the prompt we supply to Claude Code. They reflect Dagster's engineering judgment about how Python should be written, and they are battle-tested against a large, actively maintained codebase. We share them here so you can adopt or adapt them for your own workflows. The goal is simple: impart your expertise to the model so its output reflects your team's standards, not generic patterns from its training data.

The single most important rule in Dignified Python is to check conditions proactively rather than relying on exceptions for control flow. We call this Look Before You Leap (LBYL), in contrast to the Easier to Ask for Forgiveness than Permission (EAFP) pattern that many LLMs default to.

WRONG: Exception as control flow try: value = mapping[key] process(value) except KeyError: pass # CORRECT: Check first if key in mapping: value = mapping[key] process(value)LBYL makes intent explicit. The reader can see immediately what conditions are being checked and what happens in each case. EAFP obscures this by burying th