#Abandoned.

1 messages · Page 1 of 1 (latest)

loud hawk
#

Interesting

#

you want my honest take?

#

Its a good prompt

#

but

#

Your giving the AI to much responsibility

#

yeah let me give you a example

#

heres a prompt in my system this is for the Self Assessment Agent

#

that verfies the Builders work

#

I made its job clear right off the bat and then i told it how to respond what JSON format

#

its very clear

#

its very direct

#

its easy

#

its smooth

#

the crictal rules are extra precreation but the point is simple and direct

#

when you read your prompt

#

I want you to ask yourself "If i was handed this would i do well"

#

well in my system i try to make the AI's job as easy as possible

#

so

#

usally

#

I have 1 agent thats doing the work

#

and the other is actually actulzing the work

#

So it may be in your interest to add a validator or executor

#

in my system i call it the "Intent Parser Agent"

#

because its looking at the intent of the builder or self asssemnt agent and turning it into a valid tool call

#

were not burdening the Builder or the main agent with the stress of

#

Call this tool EXACTLY

#

like this

#

we leave that job to another AI

#

and also

#

your having to tell the AI alot of "what not to do"

#

this is where you need a orcheastor

#

notice im not having to tell Self Assessment a long list of what not to do

#

because im outsourceing the work to the intent agent AND i have a orcheastor thats keeping the system automatically in check

#

Your ideas are good they just need to be refined

#

hmmm I sort of understand

#

however my advice and approach comes from the end of making it simpler and easier for the AI

#

just like you would a person

#

I would but im not sure how to adapt it to my system as ive curated a very spefic envirment of how to call tools and makes choices

lofty wraith
#

Axiom: I found a missing layer. (beta)

lofty wraith
#

Axiom: A governing prompt to stop hallucination and drift. (beta)

wild wigeon
#

We should talk - we are working on this issue more broadly. But this isn't a prompt engineering problem and is located at the wrong layer. LLMs are stochastic at the generation level, even with temperature tuned to 0 you can't guarantee deterministic output.

lofty wraith
#

Axiom: Deterministic Control Over Stochastic Models