Tips and Tricks for chat sessions, Agent-Assistant use, errors (from a non-dev)

Tips & Tricks
1
  1. In the beginning of each chat session, it’s a good practice to put a small sentence in order to create a context. “In this chat session we are going to work on…. in order to …………”
  2. Small well descriptive tasks worked within one unique chat session
  3. Focusing on the result we want rather than the way this should happen
  4. In case of minor low performance execution, retry with better context
  5. In case of errors or app crashes, try just once to fix it. In case this doesn’t work rollback or create a new chat session with more analytical context.
  6. Avoid trying to fix errors as this cause more problems than it solves and also sometimes causes unwanted changes to the app
  7. Chat with Assistant, not with Agent
  8. After chating with Assistant about a new feature ask him to write the prompt so that you can give it to Agent (with copy paste) for implementation
  9. Create new features with Agent. Don’t use the additional suggestions of Agent in the beginning of a new chat
  10. Enhance, fix, change, adjust existing features with Assistant not with Agent
  11. In case Assistant can’t execute your request, try it with Agent in a new chat (but don’t overdoit)

Good luck

4 Likes
2

Very good and clear! Thanks!

3

I’ve found that before anything I go to ChatGPT/Claude, and have it create a PRD - a Product Requirement Document, with instructions to break down each part into phases and to tell me for each phase, do I use the Agent or the Assistant. Sometimes depending on the size of the phase, I revise to break it down into descriptive sub-phrases so that there isn’t any confusion

For each new phase, use a new chat.

It has worked wonderful.

Agent = building the house. The box. The main bones, support beams, etc of the app
Assistant = putting in hardwood floors, marble countertops, this years Pantone color of the year on the walls, trim, etc.

1 Like
4

I seem to get caught in this cycle of error fixings … to your point at number 6, if this can be avoided then by all means do not go down that rabbit hole. Unfortunately I’ve been trying to build the same app over the last couple of weeks. 7 complete rebuilds, different errors randomly appearing that seem just unfixable. Almost ready to give up.

5

I found that sometimes when you iterate through the errors you can identify a loop, then you say something like “I noticed an error loop as follows: when you fix X error it causes Y error, then fixing Y error causes Z error, then fixing Z error causes the original X error, repeat”, then Assistant says something like “Oh! That means yadayadayada” and then it fixes the problem.

I’d say with my approach it’s like a 50% chance of success. Probably could increase that if I was technical or had better prompts.

6

Sometimes if you call out the AI saying “we’ve already done this. Stop it” its like oh yeah haha