I'm building a startup myself and we vibe code everything, except core/complex parts where AI struggles.
My workflow:
1. Give the agent detailed task + relevant file contexts (so it doesn't waste time searching)
2. Guide it with continuous feedback when it veers off
3. Jump in manually if it's going completely wrong direction
I use Claude Code for the agent work, Cursor for manual edits.
Key insight: AI coding works as well as your architecture does. I make all architecture decisions, file structure, organization -- AI just writes the implementation logic.
Clean architecture = better AI output.
I believe This is becoming the norm because it gives startups massive velocity advantage. That's probably why your CTO is pushing it.
Career impact? As a senior who understands what AI writes, you're gaining superpowers, not losing skills. You were previously bottlenecked by implementation speed and needed other engineers. Now you can be a 10x engineer - as long as you can envision the architecture, AI helps you build it faster.
The key is knowing when to guide vs when to take over. That judgment comes from experience.
As a former senior solution architect, I have tested several tools or platforms of "vibe-coding" or "agentic software development". None of the available tools or platforms satisfied the quality level, I used to have. Number of failures was more than successes. Which is completely understandable for the "Sophisticated Prediction Algorithms" or LLMs in official terms.
The main problem is these systems do not have proper judgement mechanisms to translate user or customer requirements to executable tasks.
Then, I have ended up with https://github.com/bmad-code-org/BMAD-METHOD. This framework is a well thought approach to create very detailed definitions of the tasks that LLM based coding agents can process and execute with better outcomes.
I'm building a startup myself and we vibe code everything, except core/complex parts where AI struggles.
My workflow:
1. Give the agent detailed task + relevant file contexts (so it doesn't waste time searching)
2. Guide it with continuous feedback when it veers off
3. Jump in manually if it's going completely wrong direction
I use Claude Code for the agent work, Cursor for manual edits.
Key insight: AI coding works as well as your architecture does. I make all architecture decisions, file structure, organization -- AI just writes the implementation logic.
Clean architecture = better AI output.
I believe This is becoming the norm because it gives startups massive velocity advantage. That's probably why your CTO is pushing it.
Career impact? As a senior who understands what AI writes, you're gaining superpowers, not losing skills. You were previously bottlenecked by implementation speed and needed other engineers. Now you can be a 10x engineer - as long as you can envision the architecture, AI helps you build it faster.
The key is knowing when to guide vs when to take over. That judgment comes from experience.
Thank you for your response.
If it's ok, could you please share if you follow a certain structure when defining the detailed task for the agent.
Hi,
As a former senior solution architect, I have tested several tools or platforms of "vibe-coding" or "agentic software development". None of the available tools or platforms satisfied the quality level, I used to have. Number of failures was more than successes. Which is completely understandable for the "Sophisticated Prediction Algorithms" or LLMs in official terms.
The main problem is these systems do not have proper judgement mechanisms to translate user or customer requirements to executable tasks.
Then, I have ended up with https://github.com/bmad-code-org/BMAD-METHOD. This framework is a well thought approach to create very detailed definitions of the tasks that LLM based coding agents can process and execute with better outcomes.
To address this topic I have recently published an article you can read https://open.substack.com/pub/practicalsecurity/p/the-70-pro... . I hope the information and research for this article gives you some ammunition for your defense.
For further discussions feel free to reach out by leaving a comment to the article if you wish.
I hope things will get better for you in the startup you are contributing.
Thank you for your reply!
I never heard of BMAD-METHOD, so thank you for sharing this. I will go through it and definitely get back to you via your article.