I did a big verbal dump on what I've learned coding with AI and Cursor and made Grok organize it for me
you can find my draft below
And the overview-
you can find my draft below
And the overview-
https://twitter.com/simonw/status/1899463121338777629
My top takes:
- a code orchestrator's job is to provide the best context
- do a couple of trial one-shots with your PRD to find unclarified areas
- use the Claude models more as a toolkit vs using just one
- make sure to let llms run wild from time to time to see it's limits
- a code orchestrator's job is to provide the best context
- do a couple of trial one-shots with your PRD to find unclarified areas
- use the Claude models more as a toolkit vs using just one
- make sure to let llms run wild from time to time to see it's limits
Provide Cursor with too much context and it gets confused
Provide it with too little and it duplicates code and starts making new patterns
You should know what every file does to get the best results
Provide it with too little and it duplicates code and starts making new patterns
You should know what every file does to get the best results
https://x.com/bydallasklein/status/1897724230763991510
Don't be afraid to try out some prompts and discard the changes
When I start a new project, I'll always let Cursor run wild with my PRD (10-15 min) and see what it builds
I get a better feel for how the LLM understands what I want to build and fill in any missing gaps
When I start a new project, I'll always let Cursor run wild with my PRD (10-15 min) and see what it builds
I get a better feel for how the LLM understands what I want to build and fill in any missing gaps
A lot of people are complainging about Claude 3.7
but I think they complement 3.5 nicely
and thus- use Claude as a toolkit
instead of using one model 100% of the time
but I think they complement 3.5 nicely
and thus- use Claude as a toolkit
instead of using one model 100% of the time
https://x.com/bydallasklein/status/1898858735096734196
Don't assume you know what the models can do, push it to the limit sometimes
Try out different levels of specificity for your prompts
sometimes I'm surprised at how much it can actually do
and that I've been restricting it
LLMs continue to surprise us! Give it a chance!
Try out different levels of specificity for your prompts
sometimes I'm surprised at how much it can actually do
and that I've been restricting it
LLMs continue to surprise us! Give it a chance!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
