17 posts
Large language models are decent brainstorming partners for discovering test oracles, but they hallucinate properties and miss domain constraints.
You need to test a function where the correct output is impossible to know in advance. A route optimizer. A sentiment classifier. A physics simulation. You've…
LLM conversations contain intent, rejected alternatives, and working code. That's exactly what documentation should be.
You spent forty-five minutes with Claude designing a retry circuit. You explained the failure modes, rejected exponential backoff because it hides cascading…
Fagan inspections need four to six people and two hours to review 250 lines.
A full Fagan inspection needs a moderator, a reader, two to four inspectors, and the author. The team spends two hours reviewing roughly 250 lines of code at…
N-version programming sounds simple: run multiple implementations and pick the best answer. In practice, 'best' is a lot harder to define than 'most common.'
You have five implementations of the same function. Three return the same result. One is slightly different. One throws an exception. Which one is right? Most…
N-version programming with LLMs doesn't require multiple models.
N-version programming assumes diversity comes from different authors. With LLMs, that means different models, different providers, maybe different training…
AI coding follows a validation gradient. Backend validates in milliseconds. Web takes minutes. Mobile takes hours. The feedback loop determines where AI coding works.
An AI model coding by itself is not the interesting part. The interesting part is what happens after it generates the code. How fast can the model know if the…
AI coding works for elite teams who recover from any mess. Everyone else gets fear, failed CI, and abandoned experiments. The gap isn't the model.
Watch two teams use the same AI model and you will see two completely different outcomes. The first team prompts the model to build a screen. The output is…
Most teams try AI coding, ship code that fails QA, and quit. The problem isn't the model—it's the absence of guardrails that make AI output trustworthy.
Most teams that try AI coding follow the same arc. They start excited. The model generates a feature in minutes, and they ship it. QA catches a bug, so they…
You have type checking, linting, and architecture rules. But your deterministic stack is blind to complexity, duplication, and naming disasters. Here's the $0 fix.
Let's be honest about what most AI code pipelines actually look like right now. You generate code with Cursor or Claude Code. You run because TypeScript strict…
AI coding tools excel at generating version one. The real engineering challenge starts at version four, when the team needs to change something without breaking everything else.
Every AI coding demo follows the same arc. Someone prompts a model. A working app materializes. The audience is impressed. And they should be. The speed is…
The real measure of AI-generated code quality is not whether it works on day one. It is whether you can replace it on day thirty without rewriting everything else.
Most conversations about AI-generated code quality focus on correctness at generation time. Does the output compile? Does it pass the tests? Does it match the…
Human review is inconsistent. AI review is worse. The only scalable defense for AI-generated codebases is deterministic enforcement: rules that fail the build, not suggestions that get ignored.
The standard advice for AI-generated code is "review it carefully." That advice is correct and useless at scale. A developer reviewing AI output catches…
When AI can generate implementation, tests, and contracts from a spec, the highest-leverage human work moves upstream.
If you have been shipping with AI for more than a few weeks, you probably know this feeling. You open a pull request. The code is clean enough. The naming is…
If you want AI-generated code to survive production, do not rely on code review alone.
The most dangerous thing about AI-generated code is not that it is always wrong. It is that it is often good enough to merge. That is exactly what makes it…
The meme is funny because teams really do build against moving targets, fixed launch dates, and half-decided scope.
Call it Meteor development. If you first encountered the meme in Chinese or Japanese circles, that tracks: in Chinese writing it often appears as 陨石开发 / 隕石開發,…
A Stanford CS146S review and critique: the inaugural Fall 2025 course gets AI coding right, but the missing lesson is AI architecture and replaceability.
Stanford CS146S is CS146S: The Modern Software Developer, taught by Mihail Eric and launched as an inaugural Fall 2025 course. If you want the official…
Design by contract, property-based testing, mutation testing, and model checking were not bad ideas. They were too expertise-heavy to sustain.
Software engineering is full of ideas that look obviously correct the moment you read them. Of course contracts should define what a function is allowed to…