Creating AI personas

I’m always typically late to most things, it’s just an unfortunate habit of mine, and if i’d realised it sooner in life, i’d have made a fortune (i’m talking trading obviously :)

The same thing appears to be happening with my journey with Claude Code, because as soon as i finally start making some persona markdown files, some big video blog post comes out saying why that is actually a bad idea and can impact the model and its responses. ffs.

Regardless, i’ll carry on exploring it for now because it proved very useful for what i’ve been using it for. Specifically trying to convince one of my devs at work that some of the stuff i’m coming up with using Claude, is actually pretty good. So to front run his continuous hole picking in everything, i created a devilsadvocate persona based on said developer. Devilsadvocate does exactly that, it picks holes in everything, it is super critical and will always find a problem even if there isn’t really a problem, it will make one up lol.

Armed with Devilsadvocate, i can now prompt claude to devilsadvocate its own work, then prompt again with the feedback and ask Claude to fix the problems found. I just iterate this a few times until devilsadvocate is happy. And what do you know, just today my lead dev didn’t really have much to say when i showed him the latest thing i’d put together to help us with some code documentation we’ve been meaning to do for some time now. In fact, he went silent, which was very out of character. I think he spent most of the afternoon looking through the repo, and realising that these agents are actually getting a lot better, and it is now building small applications with unit test and proper documentation readme’s, rather than just me copying some code out of a chat window and running python scripts.

I then decided i was bored of trying to explain what i was building to a wider audience, so i got Claude to give me a summary. As usual, the summary was very good, detailed, no spelling mistakes, concise and clear… too much like an AI. So it suddenly occurred to me that the Devilsadvocate persona had done such a good job at mirroring my lead dev, that i decided to create a persona based on myself.

After a few iterations, and with me retyping out what Claude had put together, we hit the sweet spot. I can now get AI to give me a summary of any project i build in my own voice and tone, so it doesn’t look like i’m being lazy (i am) and getting AI to summarise its own work. The temptation to then start using my persona on Claude to reply to any email coming in whats tugging at me, but for now, i’m not going to open that particular can of worms.

Like i said at the start, i’m usually late to things, and i know this isn’t a new technique when working with AI, its just that i’ve never bothered before when using just a chat bot. But with an agent living on your machine, and the ability to write and save directly to your computers file system, its become a no-brainer.

Anyway, i came across this quote the other day and loved it

Someday a computer will give a wrong answer to spare someone’s feelings, and man will have invented artificial intelligence. Robert Brault, 2011