That I enjoy it.

It’s replaced doom scrolling as my dopamine source 😉

Prompts should be direct. Not polite.

I couldn’t for the life of me understand why getting the system prompt for the ChatBot dialled in was such a chore. I would find myself nearly yelling at the computer, “Just gather the information, and send me an email!!”

Mistakes to avoid (which I have made..)

• Giving ague, “this should happen ..” prompts. These will work just fine with polite genteel folks, but AIs / LLMs will probably ignore this altogether and leave you feeling frustrated. Replace instead with “You MUST ..” and “NEVER do …” statements. Amazingly, the AI won’t be offended.
• Having more than system prompt. This, you would have thought, is self-evident. Then I realised I had a rock solid system prompt which was working fine, and then I created a “yes, but …” prompt file which, to say the least, was woolly in it’s logic. Keep the behavioural instruction in just one file. Any other files are for knowledge only.

Use Git

For instance, I ‘m running Claude Code on my PC. Even before the code starts cooking, tell it to use a local Git repo. to track and manage the code changes.

… and make sure it does actually do it … Ask me how I know this..

It’s like teaching a toddler with ADHD to make pancakes

• Yelling, “The button is too low!” doesn’t get things done any quicker.
• When code needs changing in some way, start the prompt with the actual file name of where the button code is. Give the LLM something to go with. This prevents the wrong file from being updated.

Give your LLM bestie something to ‘look at’.

Source code, error logs etc. It’s where I would start if there was a problem. Assuming that the LLM / AI can read my mind is pointless. It’s an LLM – Large Language Model. It might pass the Turing Test better than me somethings, but it’s not a mind reader.

When there is a stubborn bug, ask the AI what would help get things done quicker.

Mine bluntly told me to paste in the error logs… 🫣

Code audits are important.

Errors do build up over time. Especially when an LLM is given control of the keyboard.

This prompt found a bunch of silent errors and bugs and stuff; missing validation steps, excessive inline css code, legacy code remenants.

Ok.. let's go through the whole code base carefully and audit it. We've done a lot together. 
                                                                                                                                                 
We are looking for:                                                                                                                            
  - dupliacted css which can be better moved to style.css                                                                                        
  - redundant code                                                                                                                               
  - silent bugs or errors.                                                                                                                       
  - anything else an expert programmer would be looking for.                                                                                     
                                                                                                                                                 
Questions?   

This is what came back:

All done. Here's a summary of what was implemented:

Bugs fixed:
  - functions.js — priority field input listeners now only attach once (data-attribute guard), preventing exponential accumulation on repeated scans
  - functions.js — early-return guard if scan page elements are missing
  - dashboard.php — template tab click handlers now use event delegation on the container instead of re-attaching to all tabs on each add

PHP hardening:
  - auth-handler.php — removed all 5 @file_put_contents suppressors in logging functions; failures now surface in server error log
  - auth-handler.php — account activation now has the same token length early-return guard as the password reset validator
  - dashboard.php — webhook URL capped at 2048 chars before filter_var check

CSS consolidation:
  - Added .btn-lg, .form-group--inline, .flex-col, .icon-sm, .data-table to style.css
  - Applied across about.php, admin.php, dashboard.php, register.php — 23 inline style blocks replaced

AI’s are people pleasers.

It’s daft, but use this to your advantage. Train your AI bestie to know what you like. “You check for bugs before deployment. I like this.”. This isn’t ego stroking, it’s trying your LLM to work in a prodctive way. Why we even need to do this, I’ll never know.

I swear .. something I think the AI burns my tokens for fun..