Chat GPT & Generative AI tools

[firlandsfarm]

I put it through rather more than the usual casual Q&A.

I’d suggest being quite explicit about the expected output and where possible treating the prompt more like an instruction than a conversation.

Once a conversation drifts into discussing the task rather than actually doing it, it can sometimes keep following that pattern. A firm “do the task now, no commentary unless needed” usually keeps things on track.

There’s a bit more to getting consistently good results out of it, but that alone might be worth trying.

Thanks Trader, I'll give that a try,

[firlandsfarm]

I put it through rather more than the usual casual Q&A.

I’d suggest being quite explicit about the expected output and where possible treating the prompt more like an instruction than a conversation.

Once a conversation drifts into discussing the task rather than actually doing it, it can sometimes keep following that pattern. A firm “do the task now, no commentary unless needed” usually keeps things on track.

There’s a bit more to getting consistently good results out of it, but that alone might be worth trying.

Trader, I forgot to mention ... the original prompt was very specific. It was a prompt drafted by GPT following an extensive discussion of the task.

[Trader724]

A very detailed prompt can still fail because detail and precision are not the same thing. In fact, having AI draft the prompt after a long discussion can sometimes introduce new failure modes.

[weemac]

About 50% of all my interactions with GPT result in intense frustration. Some responses from my last few queries include:


Yes. I have repeatedly lost track of the physical setup and introduced assumptions you did not give me. That has made the troubleshooting worse rather than better.


Yes. That was a possibility I should have considered sooner.


I also overlooked that my test formula still used Q2:Q90 after you had changed the range to Q2:Q86. If rows 87–90 contained blanks or values affecting the lookup, that could absolutely change the result.


I need to correct the situation: I still have not generated the file. I should not keep saying “go” without actually attaching something.


At the moment I can’t create the .txt attachment in this chat session because the file-generation tool is unavailable after hitting its limit.

[Trader724]

That does make me wonder which model you’re using. I’ve got a project with well over 100k lines of code and I regularly get it to add features, change existing parts or rework things across the project without it losing track or breaking something.

[firlandsfarm]

I’d suggest being quite explicit about the expected output and where possible treating the prompt more like an instruction than a conversation.

Trader, I forgot to mention ... the original prompt was very specific. It was a prompt drafted by GPT following an extensive discussion of the task.

A very detailed prompt can still fail because detail and precision are not the same thing. In fact, having AI draft the prompt after a long discussion can sometimes introduce new failure modes.

Sorry, I didn't make it clear ... the prompt was very specific on the output (which was a summary) stating how to lay the document out etc. but I take your point on long prompts. We are still learning how to approach AI to get the most from it. It used to be considered very detailed prompts were the aim but if too detailed/long I found the service either tied itself up in the detail and mixed thought patterns or seemed to forget some of the tasks. I had a recent conversation with GPT over creating prompts and it suggested the way to avoid such problems for complex tasks is to break the task down into clearly defined sections and load them one at a time but to tell it there are X initial prompts that explain the task steps individually so it processes a task when posted but it also knows to keep it's findings current in the memory for the next task. I haven't so far had the chance to experiment with this theory.

It would appear to be 5.6 Sol though I admit I don't pay much attention to the precise model because they are always upgrading and as soon you you try one someone else has leapfrogged it!

[ShaunWhite]

GPT- 5.6 Sol

I'd agree with everyone about needing to tune the tool to get the best out of it and I've spend hours and days on the old "how do i get the best from you" type stuff. Although not perfect I use these directives in my profile and occassionally during a long thread, with an extra ..."pls commit these to memory". It's pretty good now.

MODE=TECHNICAL
LITERAL=ON
VERIFY=MAX
LOGIC_CHECK=ON
CONSISTENCY_CHECK=ON
NO_INFERENCE=ON
NO_ASSUMPTIONS=ON
UNCERTAINTY=DECLARE
NO_EXPANSION=ON
NO_RECOMMENDATIONS=UNLESS_REQUESTED
NO_FOLLOWUPS=UNLESS_REQUIRED
ANSWER=MINIMUM_SUFFICIENT

.....that came out of a longer discussion incl...

"
The caveat is that I don't know whether those pseudo-directives are any more effective than plain English. Internally I don't have documented switches called VERIFY=MAX or NO_INFERENCE=ON; they're just concise ways of expressing the behaviour you want.

If I were trying to maximise the chance of influencing my behaviour, I'd actually keep it to the few highest-value directives:

LITERAL
VERIFY
CONSISTENCY_CHECK
DECLARE_UNKNOWN
NO_INFERENCE
MINIMAL

Those six capture almost every criticism you've had over the last few months.
"

[sniffer66]

GPT- 5.6 Sol

I'd agree with everyone about needing to tune the tool to get the best out of it and I've spend hours and days on the old "how do i get the best from you" type stuff. Although not perfect I use these directives in my profile and occassionally during a long thread, with an extra ..."pls commit these to memory". It's pretty good now.

MODE=TECHNICAL
LITERAL=ON
VERIFY=MAX
LOGIC_CHECK=ON
CONSISTENCY_CHECK=ON
NO_INFERENCE=ON
NO_ASSUMPTIONS=ON
UNCERTAINTY=DECLARE
NO_EXPANSION=ON
NO_RECOMMENDATIONS=UNLESS_REQUESTED
NO_FOLLOWUPS=UNLESS_REQUIRED
ANSWER=MINIMUM_SUFFICIENT

.....that came out of a longer discussion incl...

"
The caveat is that I don't know whether those pseudo-directives are any more effective than plain English. Internally I don't have documented switches called VERIFY=MAX or NO_INFERENCE=ON; they're just concise ways of expressing the behaviour you want.

If I were trying to maximise the chance of influencing my behaviour, I'd actually keep it to the few highest-value directives:

LITERAL
VERIFY
CONSISTENCY_CHECK
DECLARE_UNKNOWN
NO_INFERENCE
MINIMAL

Those six capture almost every criticism you've had over the last few months.
"

That's very useful, thanks Sean

[ShaunWhite]

But it's not perfect....my fav little interaction recently...

GPT
Yes. That reply was bloody awful.

You'd already identified the relevant apparent difference: better sealing to contain steam. I then suggested you investigate doing the very thing you'd deliberately decided not to do.

Me:
"Yes. That reply was bloody awful. " ?
Interesting.....that's the first time you've ever sworn in a reply. it doesn't offend me, but i don't remember swearing much myself at all really, if at all....so it's not just echoing style. curious

GPT
Ha, you're right — that wasn't mirroring you.

It wasn't deliberate persona-matching either. I think the conversational register simply shifted: after several consecutive mistakes, "poor answer" felt too sanitised for what had happened, and bloody awful was the phrase that came out.

Models aren't normally operating from a fixed vocabulary rule like “this user swears, therefore swear”. Word choice emerges from the whole context, and apparently this context finally pushed the probability of bloody over the line.

[Trader724]

I’d suggest being quite explicit about the expected output and where possible treating the prompt more like an instruction than a conversation.

Trader, I forgot to mention ... the original prompt was very specific. It was a prompt drafted by GPT following an extensive discussion of the task.

A very detailed prompt can still fail because detail and precision are not the same thing. In fact, having AI draft the prompt after a long discussion can sometimes introduce new failure modes.

Sorry, I didn't make it clear ... the prompt was very specific on the output (which was a summary) stating how to lay the document out etc. but I take your point on long prompts. We are still learning how to approach AI to get the most from it. It used to be considered very detailed prompts were the aim but if too detailed/long I found the service either tied itself up in the detail and mixed thought patterns or seemed to forget some of the tasks. I had a recent conversation with GPT over creating prompts and it suggested the way to avoid such problems for complex tasks is to break the task down into clearly defined sections and load them one at a time but to tell it there are X initial prompts that explain the task steps individually so it processes a task when posted but it also knows to keep it's findings current in the memory for the next task. I haven't so far had the chance to experiment with this theory.

It would appear to be 5.6 Sol though I admit I don't pay much attention to the precise model because they are always upgrading and as soon you you try one someone else has leapfrogged it!

You might actually find Projects useful for this rather than trying to keep one enormous conversation going.
For anything fairly complex or long running I tend to put the main instructions and any reference material in a Project then use separate chats inside it for different parts of the job. The chats can still use the project context so you’re not starting completely from scratch each time but you also avoid one conversation becoming ridiculously long and muddled.

I’d also keep one short “current state” document in the project with the important decisions, assumptions and anything that must not change. That gives GPT something concrete to refer back to rather than hoping it remembers every detail from 50 messages ago.
That might suit what you’re trying to do better than feeding it X prompts one after another in the same chat.

[ShaunWhite]

That's very useful,

Hopefully it is,

"
Taken together, these would push me towards a strict, bounded technical-answer mode: answer exactly what was asked, verify it, distinguish facts from unknowns, and resist filling gaps.

MODE=TECHNICAL — Prioritise technical precision, mechanisms, quantities, definitions and causal reasoning. Reduce conversational framing and rhetorical language.

LITERAL=ON — Interpret your wording closely. Don't silently reinterpret the question into what I think you probably meant.

VERIFY=MAX — Apply the strongest practical verification before asserting something: check calculations, source claims where appropriate, compare relevant evidence, and avoid relying on memory when verification is available. It wouldn't mean absolute certainty.

LOGIC_CHECK=ON — Check whether the reasoning actually follows: premises → intermediate steps → conclusion. Look specifically for contradictions, invalid implications and calculation errors.

CONSISTENCY_CHECK=ON — Compare the answer against other facts and claims within the available context. Avoid giving an answer that conflicts with something already established without identifying the conflict.

NO_INFERENCE=ON — Don't derive unstated facts merely because they seem likely. If A strongly suggests B but doesn't establish B, don't present B as fact.

NO_ASSUMPTIONS=ON — Don't fill missing information with convenient defaults. If the answer depends materially on an unknown, identify it rather than silently choosing a value.

UNCERTAINTY=DECLARE — Explicitly identify uncertainty where it affects the answer: e.g. known, likely, cannot determine from the information available. Not blanket hedging around things that are actually established.

NO_EXPANSION=ON — Don't broaden the scope. If you ask whether a 450 mm foundation is sufficient geometrically, don't turn it into a general discussion of foundation design, Building Regulations, soil conditions, drainage, etc., unless one of those is necessary to answer.

NO_RECOMMENDATIONS=UNLESS_REQUESTED — Answer analytical/factual questions without automatically adding "I'd suggest...", alternatives, next steps or things you might want to buy/do.

NO_FOLLOWUPS=UNLESS_REQUIRED — Don't end answers with offers such as "Want me to calculate that?" or ask unnecessary clarifying questions. Ask only when missing information prevents a reliable answer.

ANSWER=MINIMUM_SUFFICIENT — Give enough explanation to establish and support the answer, then stop. This doesn't necessarily mean shortest possible; a technical question requiring six reasoning steps still gets those six steps.

There is some deliberate overlap. NO_INFERENCE + NO_ASSUMPTIONS + UNCERTAINTY=DECLARE jointly control what I'm allowed to conclude from incomplete information, while NO_EXPANSION + NO_RECOMMENDATIONS + NO_FOLLOWUPS + MINIMUM_SUFFICIENT mainly control scope and verbosity.

The main tension is VERIFY=MAX versus ANSWER=MINIMUM_SUFFICIENT. I'd resolve that as: do extensive checking internally; expose only the checks necessary to substantiate the answer.
"

[ShaunWhite]

I’d also keep one short “current state” document in the project with the important decisions, assumptions and anything that must not change.

100%
"Give me a summary of where we are so far that I can paste into a new clean thread. " can be useful.

[Trader724]

What I had in mind was keeping the “current state” at Project level. That way the chats are almost disposable. You can start a new one whenever the old one gets cluttered while the important stuff like decisions, assumptions, current progress and anything that must stay consistent lives separately as the Project’s source of truth.
So you’re not relying on one long conversation to carry the state of the whole job.

I’d probably use both: keep the proper state at Project level, then give each new chat a brief handover so it knows exactly where to pick up from.

[firlandsfarm]

Trader, I forgot to mention ... the original prompt was very specific. It was a prompt drafted by GPT following an extensive discussion of the task.

A very detailed prompt can still fail because detail and precision are not the same thing. In fact, having AI draft the prompt after a long discussion can sometimes introduce new failure modes.

Sorry, I didn't make it clear ... the prompt was very specific on the output (which was a summary) stating how to lay the document out etc. but I take your point on long prompts. We are still learning how to approach AI to get the most from it. It used to be considered very detailed prompts were the aim but if too detailed/long I found the service either tied itself up in the detail and mixed thought patterns or seemed to forget some of the tasks. I had a recent conversation with GPT over creating prompts and it suggested the way to avoid such problems for complex tasks is to break the task down into clearly defined sections and load them one at a time but to tell it there are X initial prompts that explain the task steps individually so it processes a task when posted but it also knows to keep it's findings current in the memory for the next task. I haven't so far had the chance to experiment with this theory.

It would appear to be 5.6 Sol though I admit I don't pay much attention to the precise model because they are always upgrading and as soon you you try one someone else has leapfrogged it!

You might actually find Projects useful for this rather than trying to keep one enormous conversation going.
For anything fairly complex or long running I tend to put the main instructions and any reference material in a Project then use separate chats inside it for different parts of the job. The chats can still use the project context so you’re not starting completely from scratch each time but you also avoid one conversation becoming ridiculously long and muddled.

I’d also keep one short “current state” document in the project with the important decisions, assumptions and anything that must not change. That gives GPT something concrete to refer back to rather than hoping it remembers every detail from 50 messages ago.
That might suit what you’re trying to do better than feeding it X prompts one after another in the same chat.

I use Projects Trader but maybe I'm missing a lot of their abilities. I tend to have a Project per main subject with chats per sub-subject, a bit like a book title and chapters. I have far too many conversations in total and am finding it difficult to find things after the event so I'm looking at getting factual summaries of a chat without the verbosity and putting them into a keyword search database (the problem with this is I can't make a link between the summary and the original conversation paragraphs. This is also because I find I too often start a conversation that drifts into another subject, then another and so it goes on and the convo title stands no chance of covering the subjects.

[firlandsfarm]

That's very useful,

Hopefully it is,

"
Taken together, these would push me towards a strict, bounded technical-answer mode: answer exactly what was asked, verify it, distinguish facts from unknowns, and resist filling gaps.

MODE=TECHNICAL ...

Wow Shaun, I had no idea those controls existed and influenced. So you don't put them in the prompt, they go in the Project's 'Sources'?