Hacker News new | ask | show | jobs
by ok1984 21 days ago
My process is becoming as follows:

1) I always start with a clean code base (ie. no uncommitted changes)

2) I have a text file called prompts in which I write my first prompt for my desired change, in the initial prompt I always explicitly ask to NOT generate unit tests.

3) Copy paste the prompt to the cli

4) While the LLM is working on my prompt I go back to my prompts file and start preparing the next prompts that will build on top of the previous one

5) Once the agent is done I review every single line of code and either manually fix what I don’t like or prompt minor changes until I reach my desired result.

6) commit locally the change

7) copy paste the next major prompt from my prompts file

8) Repeat steps 4-8 until I fully implement my major change

9) Once I have my major change I prompt for adding unit tests and related documentation

10) Once my unit tests are all green I execute additional manual tests if required, squash all local commits and push for pear review

Following these steps I can work fully focused and let me emphasise again on the fact that even if the majority of code is written by the LLM I still review and read every single line of code so that when LLM companies decide to fuck us or when shit happens in production I know where to look as if I wrote the code myself like the old days

All of this works for me because I know very well the programming language and the specific domain I am working in thanks to the past experience collected in past years before LLMs.

2 comments

Thank you for your suggestion. Going straight to the next prompt while the agent is still working will probably help me stay in flow state longer. However I'll still have to context switch when checking the code. Nonetheless, I'll experiment with this, thanks!
Yes, its like when you are chatting with somebody and due to the async process somehow you start chatting about two things in parallel (maybe it’s only me),

Anyhow, I think that you can keep in the flow as long as the successive prompts are still related to the major change if not then I guess you deserve a stop and pause before moving on with the next task.

Quite similar, but I ask for "/grilling" skill in every prompt so the agent aligns to my decissions before it drifts into own. And I actually ask to make tests before any implementation and let them all fail.
TDD advocates would be proud of you ;) differently from your approach I still prefer to make sure the model code is inline with my expectations and only then ask for unit tests, only exception is when I am dealing with a bug then the failing unit test is a MUST.