Escaping the AI Coding Death Spiral
Why you should stop arguing with your LLM and start getting the code you want.
The Single Most Effective Solution
“When the LLM does something wrong, don’t continue the conversation. Instead, rewrite your original question with extra requirements, and then let the LLM regenerate the response.”
The “Aggressive” Reality of AI Coding
You’re not alone. The promise of effortless automation often meets the reality of high-speed frustration. This section outlines the common pain points developers experience when an AI coding session goes wrong.
“Yes, finally someone who’s making my experiences with ai coding. It’s not only Gemini. Man, I was never so fast, but also so aggressive through these models. Be careful with the agents.”
“The other annoying thing is when it changes all the comments for that one change. Like all the code has to explain the one line it added?”
“…outdated comments from LLMs piling up eternally”
“WHAT ?!”
How to Fix It: “Regenerate” vs. “Repair”
The key is to manage the AI’s context. Arguing with bad code pollutes the context window. Editing your original prompt and regenerating gives the AI a clean slate. This interactive diagram shows the two paths.
DON’T DO THIS: The “Repair” Death Spiral
Continuing the conversation pollutes the context.
Context is polluted. You are in the Death Spiral.
DO THIS: The “Regenerate” Clean Slate
Editing the original prompt gives a clean slate.
You get good, clean code.
The New Workflow: Best Practices
Beyond the “Igor Method,” the developer community suggests a more cautious and specific approach to AI coding. Here are the key takeaways for a less frustrating experience.
-
1
Keep it Small
Focus on “small snippets” and “small scripts.” Reliability drops significantly as complexity increases.
-
2
Be Specific
Don’t rely on “plain English.” Be “quite specific to make it not a hassle.” Use technical constraints in your prompts.
-
3
Use the “Igor Method”
If the AI errors, **do not reply.** Go back, edit your original prompt with more detail, and regenerate.
-
4
Verify Everything
“Always double check. It’s just a tool.” Assume the code is broken until you prove it works. Beware of AI “gaslighting.”
As Tommy commented on the thread: “Finally someone said it.”
Read the long thread : https://x.com/burkov/status/1992537151821451640