#Strictly adhere to stepped instructions

1 messages · Page 1 of 1 (latest)

fiery moat
#

So I've come to the conclusion that the way you word your instructions heavily influences the way the GPT adheres to it. Has anyone found a foolproof way of making the GPT follow a stepped approach to finding an answer to a question every time ?

So in other words:

Step 1 : Browse <URL> for answer.
Step 2: Look at document DEF
Step 3: Use your trained knowledge

But do it exactly in that order every time a question is asked?

I find mine skips step 1 all the time and just goes straight to step 2 & 3

Thanks

knotty frigate
#

Perhaps try telling it 'Step 1 : Browse <URL> for more details about the question.'

When you tell it to browse 'for answer' it makes a judgement call, does it think that url would provide the answer?

The way the user makes the input also affects this. If there's no clear reason to browse, it's harder for the model to take that action.

#

Also remember that to the model, the input is 'closer' than the instructions. The instructions guide, but the input is what the actual current user cares about and wants.

There's all the training under your instructions too, guiding the model towards how to answer that user's input. So what you tell it to do has to fit all the other stuff it has as goals and methods to answer that user's specific request

fair adder
knotty frigate
# fair adder Could you explain more your last paragraph? Also, have you had much luck with h...

My last paragraph:

Everything you tell the AI that either contradicts or is unclear is going to cause problems.

You need to give clear and specific guidance.

For example, if that's your literal instructions, it has no idea what to do. It would already infer that it should use its trained knowledge - but that includes both training data and the stuff uploaded, as you phrase it there.

You can fix that by defining 'trained knowledge' but I don't see a definition. Maybe it's in the document DEF but you don't provide that info.

Your exact wording of all of it can matter greatly to how the AI interprets and responds.

To continue, I'd urge asking 3.5 (because you get so many uses with 3.5, and can quickly explore everything without long waits) to interpret parts of you instructions, initially small parts, and make sure what you think you are telling the model is what it thinks you're telling it, and start fixing inconsistencies in language used/interpreted and look for assumptions you need to actually explain.

For example, here's what the model thinks your 'use your trained knowledge' means:

image.png
fair adder
knotty frigate
#

And that alone can help explain why your GPT may be skipping the look up the URL because it reads all the instructions and tries to understand what you're actually saying and then follow them.

Any time you don't get the intended result you should check for 2 things:

  1. Did you specifically and precisely explain what you wanted it to do, in a way that makes sense to the model to mean only and exactly what you want done. If that's a yes... then check for:

  2. Can the model actually do what you want done. Sometimes there's a conflict with something else, like safety training or with the model's actual ability to handle information (Like it's pretty bad with math and spatial sense compared to a lot else, especially if code interpreter is not used)

knotty frigate
# fair adder Oh, I'm not OP haha

Okay. But that's my explanation of my last paragraph, be you OP or not. My last paragraph was in answer to the OP. So I used that to explain it.

fair adder
#

This is still helpful information though

#

I'm working on having it read xlsx files and conver to sqlite files

knotty frigate
fair adder
#

Also working on a coding project customgpt to walk people through various complex coding projects as a tutor/senior dev mentor

knotty frigate
# fair adder I'm working on having it read xlsx files and conver to sqlite files

I'm not a programmer, and don't know about those file types either if they're nothing about programming - but if I was trying to do this, I'd start with a short and clear xlsx file and have it be converted to sqlite file.

Then I'd check this short, clear file output for any issues or problems. I'd research then discuss the problem or issue with the model, and try to see if this was something I could fix with my language use....

Which happens and can be done even with LaTeX formating which the 3.5 and 4 models can use in their output for us to see, but for whatever reason they often use a formating style that doesn't work for their environment (and the model isn't shown the output, so it can't self-correct). So we sometimes (originally was nearly always) get crazy broken formatting instead of neatly rendered, beautiful LaTeX to see.

The fix is for us to include a reminder of how to follow the LaTeX formatting rules within our prompt, and then no problems. But initially that wasn't understood by any of us, so we thought the model couldn't do it and it didn't have a fix ...

So that's an example of a problem that you could observe, discuss, and find a fix for. Other problems you might find the model as-is cannot adjust for and correct, it is 'forced' to follow the instruction one way, and nothing you say is going to change that, it's outside of the model's ability to obey.

An example of that is getting it to tell a story or a part of a story that doesn't have a summary/conclusion at the end.

#

So, if you start with the basic short simple 'example' types and find that it can convert the really easy ones between xlsx to sqlite, then you step up the difficulty, until you find where it is failing.

Then you figure out why -

is it length? Maybe you can chunk the task.

Is it a particular detail in the code being mis-handled? Instructions might fix that.

And so on.

finite vault
#

I've been experimenting with XML and Jinja syntax in the custom instructions. This allows me to be very structured with a whole schema, as well as dynamic with macros and so on in a response template. Maybe you could try something similar? Use some coding concepts to keep your prompts structured

fair adder
fair adder
finite vault
slim drift