AI-generated code lacks common sense

TLDR 🙇:

AI-generated code lacks common sense, especially senior-level. Examples of missing behavior:

• Before providing a solution, ask whether the problem should be solved at all.
• Know when to ask for more context if it is needed to give a better solution.
• Research if there are existing utilities or dependencies that can be reused.
• Consider tradeoffs before adding a new dependency. Consider dev vs runtime deps, security, licensing, and bundle size.

In addition, it lacks elegancy. The LLM-generated code might work (or might be subtly wrong), but it will rarely be the most optimal, performant, maintainable, or pretty code. It will write more code than needed, while failing to handle edge cases.

What you can do about these limitations

You can improve the output with prompts and context, to a point.

If asking the model to "Critique this approach before writing code", the model becomes too pushy and comes up with trivial objections.

If asking for "use 80/20 best judgement", Gemini starts quoting "80/20" to justify every poor decision it makes.

I also added "senior level" and "type 1 civilization" to the personal prompt, and Gemini started quoting them even when not appropriate:

Use Option 1 (import type). It allows you to maintain the "Class" source of truth (which Web Components require) but strictly prevents accidental bundling of the implementation in consumer code—keeping your bundles slim for that Type 1 civilization efficiency.

More examples:

• Node's fs watch has a recursive: true option. Till 3 years ago, that option was not available on Linux. The LLM used that as a reason to suggest continuing to use chokidar. The limitation was resolved 3 years ago, but the model didn't pick that up.
• Vite 8 is swapping Rollup for Rolldown (Rust-powered). To reduce cross-language communication overhead, plugin hook filters were introduced as a recommended optimization. Rollup also added support a year ago. The model fails to use plugin hook filters unless explicitly prompted. As an engineer, you still need to follow the latest ecosystem developments, because the model won't do so for you.
• TypeScript has a low-level compiler API. The docs for it are virtually non-existent - the source code is the only "doc" (like the 50k loc checker.ts). 99% of the users of TypeScript just use the language server (in VSCode) and the CLI (as part of the build). There are relatively few people writing frameworks and extractors that directly use TypeScript's APIs. For tasks like "transform code with a source map", you need access to undocumented APIs (they don't appear in public types or docs). The LLM is helpless for such low-level tasks. It can't tell you why const a = "a"; and const a = "a" as const; have the same type 1 but different type checking behavior when used in class A { property = a }.

If you are writing React+Tailwind front ends, you are in the golden zone of "a ton of training data". For anyone building a framework using low-level APIs, or using an in-house compiler, the model performance degrades sharply.

Better prompts and context can only go so far. We still need models to become smarter, faster, and more up-to-date in their training.

What about vibe coding?

There are too many articles online about successfully vibe coding applications using armies of agents running in parallel. However, those posts describe a limited use case:

• They create greenfield, small apps. Often for personal use or throwaway virality.
• The code maintainability, security, and efficiency are not mentioned at all. Some pride themselves on not even looking at the output code. Or they hope a smarter LLM will come later and fix the tech debt.

Non-code performance is better

I noticed the lack of common sense and elegance in the code output because I have some experience in coding. I don't have much experience in medicine, taxes, law, finance, or social rules. The performance of models on these non-code tasks seems more impressive to me.

It is likely that my skills in these areas are already below the LLM's level, so I fail to find issues in the LLM's output. If an expert in law reads the same output, they will find it similarly unimpressive. This is called Gell-Mann Amnesia effect.

Additional constraints:

• In law and medicine, the output can be "plausible" - hallucinations are hard to spot. A bad health advice may become visible two decades later. In coding, if the syntax is wrong or a non-existent API is used, compiler errors immediately.
• Law and medicine involve a lot of "Find the precedent," "Summarize the symptoms". LLMs are very good at text summarization.
• Law and medicine require a safety bias (hedge, cite sources). Whereas when coding, defensive programming is a code smell (handling non-existent cases, relying on runtime type checking).

My personal workflow

The "yolo vibe coding" does not produce satisfactory results for me. I found more limited uses for LLMs:

Use GitHub Copilot for single-line autocomplete and next edit predictions. For small edits, it does very well at looking at surrounding code to infer your intent and desired code style.

Use GitHub Copilot Chat for quick explanation of errors, arkane shell syntax, and vibe coding throwaway scripts.

Use Gemini 3 Pro or GPT-5.2 for researching available options, getting feedback, and brainstorming. My favorite use case: write a several-page-long description of the problem, the context, and the solutions I am considering. The very act of writing down solidifies my thoughts. Then, send this to the LLM asking for "feedback, suggestions, pros/cons analysis". It may come up with a new solution, or give more reasons to go with one of my proposed solutions.

The context window of Gemini is impressive. I downloaded my journal for the entire year, and Gemini created a summary of yearly accomplishments.

Use Gemini in Chrome for a quick summary of a webpage to decide if I want to invest time into reading it. Most articles aren't worth reading.

Use Text Hoarder (an extension I built) to save purified articles as .md files. Once a year, I run Kokoro TTS on those .md files to convert them to speech. This way I canlisten to these articles at high speed while walking. I save 3k articles per year this way, which take 4 days to filter through and 2 months to listen to.

State of the art performance per category

From my personal experience, this is the state of the art of LLMs in various domains:

Coding:

• Syntax: LLMs are superhuman - they know every language in the world.
• Elegancy: coding bootcamp level.
• Common sense: reasoning models with sufficient context can achieve passable results. Though I still see models fixing a bug by deleting the failing test.

Web browsing: struggles with simple tasks like scrolling and pop-up ads. For reliability, you have to use the API directly. However, many pages don't have APIs, and each API is different - it gets complex fast. A task as simple as "summarize emails in my inbox" is tricky because you need a way to inject email content & context into the model (on the fly or async), and render the summary on top of the existing email client's UI.

World model: smart undergraduate. It read everything; experienced nothing.

Social & Emotional intelligence: superhuman? Infinite patience, RLHF-trained politeness, de-escalation, and empathy.

Speech generation: superhuman. You can generate perfect speech on-device using Kokoro. State-of-the-art commercial models go beyond perfect speech and are improving emotions, accents, and pacing.

Speech recognition: Better than human on "clean" data (audiobooks). Worse than human on real-world data due to lack of context - if the speaker's name is on the slide, a human listener will connect the dots and recognize that name in speech.

Would love to hear your thoughts on this. How does your experience with LLMs compare?

Comments

[Serhii Paiiuk]

Speech recognition: Better than human on "clean" data (audiobooks). Worse than human on real-world data due to lack of context

I wanna add that AI Speech recognition likely to fail when you say one word, even with simple word like "yes", it could get is as "yet", but saying multiple word like "Yes, it is" improves the accuracy.
Multi modal speech recognition is even worse, it confuses polish/ukrainian/russian.