My other article on this topic advocates for using code over specs :)
It’s ok to talk to the agent in code. Or create examples for it to follow.
https://softwaredoug.com/blog/2026/07/04/write-code-not-spec...
back
3 comments
> It’s ok to talk to the agent in code. Or create examples for it to follow
If you're doing this, then why not just write the effing code?
"Hey Claude, can you please make a full matrix of tests according to the following pattern and parameters? Thanks." That's 15-20 minutes of work that can run in the background while I do something else.
Exactly this. Spending a few hours crafting and debugging a test you trust save hours of pointless agent time.
What I mean by "talk to the agent in code" really is to code. Establish patterns. Then have the agent use those patterns as reference for other areas of the project.
What I miss in Claude Code is code completion at prompt line.. ah wait it gets back to IDE and Copilot.
> My other article on this topic advocates for using code over specs :)
Code is executable. Specs are not.
Prompt is executable (interpreted) code with LLM agent harnes now, just the syntax is fuzzy.
"Please print 'hello world'"
Oh no.
It's IRP.