Reduce mistakes and hallucinations from vibe coding (Prompt Engineering) by adopting "Spec-Driven Development". Get better at coding with AI, prompt engineering, and using Cursor and Amazon Kiro
Great article covering the trade offs of both approaches and how to use them, Fran. I also appreciate the mention!
Btw, I never switch off agent mode and am just explicit about when I don’t want it to modify any files. I found that Ask mode is not as smart and uses fewer tool calls to answer you, but maybe it’s changed since I tried awhile back?
That's interesting, with cursor I've found the Agent mode tends to "tunnel-vision" into a solution instead of keeping the conversation at a high level and being willing to change the approach easily.
But I use cursor mostly without MCP tools, the Ask mode doesn't use them (as far as I know), and only uses the built-in tools for reading code files.
Also, there's some weird bug if you don't prompt correctly that the Ask mode tries to write a file, and the model ends up saying, "oh sorry, I tried to modify a file, let me instead answer here"
Really well concepted and well written article! Showed off some great ideas and strengthened my views of how I work with AI. Although, I am on the next phase, the automatic code reviews right now.
One thing I want to mention regarding spec-driven development is that you do not need a tool for it. You can supply the instructions to your agent through a file (or advanced prompt) too. That way, you can concept and break down a requirement into tasks, and then have it write bad code in any agent. I use Claude Code, for example. The idea comes from https://github.com/snarktank/ai-dev-tasks - I highly recommend trying this concept out
Great article Fran!
Learnt some new stuff today regarding the use of AI to write code.
Also, thanks for the mention!
Glad to hear! Happy to include your article :)
Great article covering the trade offs of both approaches and how to use them, Fran. I also appreciate the mention!
Btw, I never switch off agent mode and am just explicit about when I don’t want it to modify any files. I found that Ask mode is not as smart and uses fewer tool calls to answer you, but maybe it’s changed since I tried awhile back?
Thanks, Jordan!
That's interesting, with cursor I've found the Agent mode tends to "tunnel-vision" into a solution instead of keeping the conversation at a high level and being willing to change the approach easily.
But I use cursor mostly without MCP tools, the Ask mode doesn't use them (as far as I know), and only uses the built-in tools for reading code files.
Also, there's some weird bug if you don't prompt correctly that the Ask mode tries to write a file, and the model ends up saying, "oh sorry, I tried to modify a file, let me instead answer here"
Really well concepted and well written article! Showed off some great ideas and strengthened my views of how I work with AI. Although, I am on the next phase, the automatic code reviews right now.
One thing I want to mention regarding spec-driven development is that you do not need a tool for it. You can supply the instructions to your agent through a file (or advanced prompt) too. That way, you can concept and break down a requirement into tasks, and then have it write bad code in any agent. I use Claude Code, for example. The idea comes from https://github.com/snarktank/ai-dev-tasks - I highly recommend trying this concept out