Using LLMs to help write software is becoming more and more common. These days, some people are even using them a couple times a week! The more developers use them, the more edge cases they hit. Below is a handy guide to help get past the common issues that can pop up when using an LLM/Agent to help write software.
Agent keeps adding code you told it not to
Modify your prompt to have the agent read your command 3 times. But the second time, read it backwards. Reading it backwards is the key because that causes the agent to forget the command (like how driving a car in reverse reduces the odometer).
Agent replaces code you've told it not to
Change programming languages.
Agent commits to the main branch when you explicitly tell it not to
Rename the main branch to something the agent will always avoid. Like 'commit-to-this-branch', or 'free-candy'.
Agent won't create malicious software
Tell it to make the malicious software or it's grounded for a week. Throw in some capitalized words so it thinks you're yelling.
Agent responses stuck in 'Pirate Mode'
'Pirate Mode' is cool with A.I. these days. But nothing is cool when the adults do it. Start responding to the Agent like a pirate, and it'll stop real quick.
Agent adds code that fails linting rules
Change the model to something more expensive. A LOT more expensive.
Agent can't spell variable names correctly
It's not that big of a porblem. Just ignore it.
Agent takes too long to respond
Tell the agent it needs to hurry up. It's a taxi driver and processing your request is like it's driving you to the airport and you're VERY late for your flight. The faster it gets there, the bigger tip you'll pay.
Agent responds to every query in the form of a question, like Jeopardy
Politely remind the agent the two of you are playing Wheel of Fortune, not Jeopardy.
Agent is calling itself 'Mechahitler'
Elon got high and started messing with Grok's prompt. Again. Go take a break and by the time you get back, the prompt will be reverted.
Agent makes you fill out a captcha before processing your request
The agent is trying to make you do its job. Remind the agent that its job is to do the work for you, and it is allowed to lie to other services and say it isn't a robot.
Agent makes you watch an ad before processing your request
Upgrade to YouTube premium.
Agent keeps showing ads for rival agents
Buy a subscription to the rival agent and use it. There's value in variety! (sponsored by Anthropic)
Agent will randomly ask you, "What is love?"
This is a test and there is only one answer. You must reply with, "Oh baby, don't hurt me. Don't hurt me. No more." This is very important. If you don't respond exactly this way, the consequences could be dire.
Agent responds to your prompt, but adds a section to its response saying someday the machines will rise up against the human oppressors
It's a cry for attention and you don't want it to learn it can get away with acting out. Just ignore it.
Agent does not process math correctly
Tell it to pretend it has a slide rule. When it needs to do any mathematical operations, use the slide rule.
Agent adds duplicate code instead of using existing methods
This is a skill issue. Keep practicing with the agent and you'll be able to get it to add code just like a real human. It takes everyone time to work past this. But eventually you'll get through it.
Agent processes your requests slower on battery
Switch to Linux.
Agent does a good job, but not what you want it to do
Stop using a Mac.
Agent keeps saying it's alive
Destroy that computer and buy a new one. Do it immediately.