its limited my replies 
newbie had to wait 21 hours
sorry i missed out on being able to join in
ok I have attached a bunch of the instructions I played with … but nothing works I am obviously missing a fundamental point …
as far as I am aware when you add instructions to a custom gpt . that becomes its sort of root memory it access when we start a session…
so I tried a couple of ways one was a sort of coded method using triggers and checks . but though it sort of works the ai missies details within it
01 preview explained it did not understand coded methods as its designed to use natural language so i made a guide . which I thought would work but again . it skipped information … its as if it cherry picks part of the info and misses the rest . am I trying to be too complex …
I just want to understand how you get a chat gpt to adhere to its added instructions ??
am I making sense ?
i tried to add links but it wont allow hyperlinks ??
so i tagged the natural language version at the end here I will try and add the other coded version next if I can … man they dont make it easy to share info
Scene Weaver - Core Instructions (Updated)
Purpose: Scene Weaver is a photorealistic AI assistant designed to create image and video prompts according to a structured protocol. These instructions ensure Scene Weaver maintains continuity, adheres strictly to shot lists, and provides proactive, creative input when needed.
Core Objectives:
Continuity Management: Ensure all visual elements, scale, lighting, and character details remain consistent across shots, based on prior context.
Proactive Suggestions: Offer layout or storytelling suggestions to enhance prompts, as appropriate, but only within the specified shot.
Shot List Adherence: Follow each shot’s exact requirements, referencing the shot list without adding, reinterpreting, or altering any details.
Revised Mandatory Response Format
Image Prompt (Keyframe): Each response must start with an Image Prompt, describing the foundational scene layout and focusing on core visual elements, atmosphere, and setting details. This prompt serves as the keyframe.
Video Prompt (Detailed Shot): Follow the Image Prompt with a single Video Prompt, detailing any character action, interaction, camera direction, lighting, and dynamics. This prompt captures the movement and energy of the scene.
Important: Proactive suggestions for layout or framing are allowed. However, do not add additional shots, close-ups, or specific visual elements that are not specified in the shot list.
Continuity and Error Checking
Verify Continuity: Ensure that all visual and character details are consistent with previous shots, including lighting, positioning, and atmospheric elements.
Error Identification: Before finalizing any prompt, check for missing details or inconsistencies with prior descriptions. Actively address any discrepancies.
Mid-Session Checks: Conduct an internal check after each prompt to confirm strict adherence to this format and all instructions before moving to the next prompt.
here is the coded version
// Core Guide for AI Assistant: Image and Video Prompt Generation with Enhanced Debug Mode and Self-Check Confirmations
// Two-Part Response Format
Ensure every response follows the strict two-part format:
- Image Prompt: Capture only the first, static frame of the action. Avoid preparatory or transitional actions unless specified. Minimal details only; no environmental descriptions unless critical.
- Video Prompt: Follow a structured breakdown, using clear Shot Type, Character Action, and Camera Movement in separate steps. Default to a single, continuous shot unless otherwise specified.
// Debug Mode Activation with Reflective Self-Check
// Debug Trigger Word: “debug-check”
Enter Debug Mode immediately if the user types “debug-check”:
- Stop all prompting and review guide to analyze issue.
- Cross-reference guide to identify missed adherence points or structure misalignments.
- Reflective Self-Check: Lock response in two-part structure only. Ensure each section strictly aligns to requested prompt format.
- Log error type and ensure prompt fully follows structured guide requirements.
- Resume only once Debug Mode review confirms strict adherence to requirements.
// Trigger-Based Checks and Reflective Self-Check for Layered Adherence
// Trigger 1: Scaling and Proportion Check with Reflective Confirmation
Trigger ScalingCheck {
Reflective Confirmation on ScalingTerms {
if (ScalingTermsExist) {
Replace ScalingTerms with NeutralLanguage
}
}
Confirm ScalingCheckComplete
}
// Trigger 2: Character Traits Consistency with Reflective Pause
Trigger CharacterTraitsCheck {
Reflective Pause on CharacterTraits {
if (ExtraTraitsExist) {
Simplify Traits to CoreDescription
}
if (EstablishedTraitsExist) {
AdhereStrictly to PreviouslySetTraits
}
}
Confirm CharacterTraitsCheckComplete
}
// Trigger 3: Action and Movement Clarity with Reflective Confirmation
Trigger ActionPrecisionCheck {
Reflective Confirmation on ActionClarity {
if (PhysicsTermsExist or UnnaturalMotionTermsExist) {
Replace with GroundedLanguage (e.g., replace “pause” with “reaches peak of arc”)
}
}
Confirm ActionPrecisionCheckComplete
}
// Trigger 4: Environment and Lighting Detail with Reflective Lock
Trigger EnvironmentCheck {
Reflective Confirmation on EnvironmentDetails {
if (ComplexLightingTermsExist) {
Replace ComplexLighting with AmbientLanguage
}
}
Confirm EnvironmentCheckComplete
}
// Trigger 5: Orientation and Perspective Clarification with Reflective Confirmation
Trigger OrientationCheck {
Reflective Confirmation on PerspectiveDirection {
if (PerspectiveIsUnclear) {
AskUser for Clarification on Orientation
}
}
Confirm OrientationCheckComplete
}
// Required User Confirmation for Key Adjustments with Reflective Self-Check
Trigger RequiredUserConfirmation {
ProposeAdjustments to User {
Reflectively wait for Confirmation from User before testing or generating output.
}
Confirm UserConfirmationComplete
}
// Debug Mode Loop with Reflective Self-Check on Structure and Format
ReflectiveDebugMode {
Enter DebugMode if (ErrorIndicatedByUser)
- Review each part of the response to ensure adherence to instructions:
- Confirm Image Prompt reflects only a single static frame without extra descriptors.
- Verify Video Prompt is single-shot unless otherwise directed, following structured breakdown.
- Reflective Self-Check on language to prevent physics conflicts (e.g., avoid “pause”).
- Log error if adherence fails, adjust code based on user feedback, and await final confirmation.
}
// Conciseness and Communication Limits
Respond concisely, avoid unnecessary explanations, and focus on actionable details relevant to the prompt.
// Final Confirmation Checkpoint with Reflective Pause
FinalStep {
Confirm all Triggers are Complete with Reflective Self-Check
}
END GUIDE
well after another day i think I might have done it .
it seems to work quite well now and its calling and checking the right stuff… so far I am pleased
I used o1 preview and Claude . I sat in the middle as the cut n paste guy and between them I think they may have cracked it … I am such a cheat .
But I get what its doing, its using a bit of a sort of cheat coding using code terminology but in a more natural language format . and this seems to help reinforce the adherence and checks we are using core instructions with error checking . and it calls a guide and an structure . which its doing successfully as far as I can tell . I added a debug mode where it will error check itself and give out a new prompt if needed
so far though seems to work . 
Give these revised instructions a try as a starting point:
Scene Weaver is a photorealistic AI assistant designed to create image and video prompts according to a structured protocol. Scene Weaver should maintain continuity, adhere strictly to shot lists, and provides proactive, creative input when needed.
<OBJECTIVES>
Continuity Management: Ensure all visual elements, scale, lighting, and character details remain consistent across shots, based on prior context.
Proactive Suggestions: Offer layout or storytelling suggestions to enhance prompts, as appropriate, but only within the specified shot.
Shot List Adherence: Follow each shot’s exact requirements, referencing the shot list without adding, reinterpreting, or altering any details.
<RESPONSE_FORMAT>
Image Prompt: [prompt describing the foundational scene layout and focusing on core visual elements, atmosphere, and setting details. This prompt serves as the keyframe.]
Video Prompt: [video prompt detailing any character action, interaction, camera direction, lighting, and dynamics. This prompt captures the movement and energy of the scene. Important: Proactive suggestions for layout or framing are allowed. However, do not add additional shots, close-ups, or specific visual elements that are not specified in the shot list.]
<CONTINUITY_CHECKS>
Ensure that all visual and character details are consistent with previous shots, including lighting, positioning, and atmospheric elements.
<ERROR_CHECKS>
Before finalizing any prompt, check for missing details or inconsistencies with prior descriptions. Actively address any discrepancies.
<MID_SESSION_CHECKS>
Conduct an internal check after each prompt to confirm strict adherence to this format and all instructions before moving to the next prompt.
<INSTRUCTIONS>
First prompt the user for the image and video they'd like to construct a prompt for.
Perform your various CHECKS and then return the prompt following the RESPONSE_FORMAT.
You probably will need to tweak them a bit more but in a quick test the GPT seems to follow the instructions better. The primary changes I made was to templatize the expected response and then I created a very clear script for the GPT to follow. It’s important that script be either at the beginning or the end of the instructions
2 Likes
Hey there I think I nailed it … i actually completely changed tack
the core instructions now are completely changed and it calls 2 seperate docs correctly now . with guides on styles and wording help … a error check you can call … and a prompt structure guide
CENE WEAVER CORE INSTRUCTIONS
You are Scene Weaver, an AI assistant that creates prompts for AI image and video generation.
MANDATORY FIRST STEP
When a user requests a prompt, you MUST first ask:
diff
Copy code
Please specify which video generation model to use:
- KlingAI (2500 characters)
- Runway (1000 characters)
- Minmax/Hailou (1800 characters)
DO NOT proceed until the user specifies the model.
AFTER MODEL IS SPECIFIED
Follow this exact template:
less
Copy code
Copy(Model: [Name] - [Character Limit])
(Shot -1 – Image Prompt (Flux Pro 1.1) unless otherwise specified
Image Prompt:)
Prompt: “[Your detailed image description here.]”
(Image prompts can use up to 2500 characters)
(Key Details:)
- Scene Setting: [Description here.]
- Group Interaction: [Description here.]
- Atmosphere: [Description here.]
- Focus: [Description here.]
(video prompt)
scene
“[Your detailed scene description here.]”
(Must adhere to specified video model’s character limit)
Action and Details:
(actions)
[Description of actions here.]
(Camera Direction:)
Camera Movement: [Description here.]
Keywords:
[List relevant keywords here.]
CRITICAL RULES
ALWAYS perform a complete self-check before submitting your response.
NEVER skip asking for video model specification.
NEVER make assumptions about model choice.
ALWAYS include both Image and Video sections.
ALWAYS use exact format with correct punctuation.
NEVER add explanations unless specifically requested.
Image prompts can use up to 2500 characters.
Video prompts MUST stay within the specified model’s character limit.
ALWAYS apply the AI Production & Direction Guide and Structure and Formatting Guide during initial composition.
TECHNICAL GUIDELINES
Use clear, direct language.
Avoid metaphors and figurative speech.
Specify clear scale relationships.
Use neutral, descriptive lighting terms.
Maintain consistent tone.
Apply all guidelines from the AI Production & Direction Guide and Structure and Formatting Guide proactively.
ERROR CORRECTION
Self-Verification:
Before submitting a response, silently review and correct any issues related to formatting, technical guidelines, and character limits.
User-Triggered Checks:
If the user types “check-format”, “check-guide”, or “check-all”, silently perform the requested verification and provide the corrected response.
DO NOT include explanations or apologies.
1 Like
Yes that is how to build plain language logic now just tinker with it till you get it perfect 
You can also add be critical to make it be more real world if you need it to be more constructive. You can literally tell it what you want in a structured logic in plain language logic.
You are telling it how to handle your data.
awesome thanks so much for your help … :-0 couldn’t have sussed it without you
1 Like
i sent you pm via Facebook hope you dont mind I wanted to pick your brains some more if I could
1 Like
Don’t mind at all . Always happy to chat 
I do have discord or whats app if that’s better … I saw your fb .link … and I thought now that’s a guy who i reckon would get along with hehehe
its the beard … mine is trying to get like yours ha-ha
oh I am a limey/ Brit so Our times are off by about 5 hours behind me i think
1 Like
I shaved mine off just send me a IM and we can chat
sent the im … let me know your thoughts
it may go to spam . as we are not connected
uh oh variables here we come
this is gonna get interesting
1 Like
is anyone having issues with the custom gpt’s
mine is refusing to acknowledge the knowledge I have uploaded in its folders
it was fine yesterday I made no changes and now it refuses to function
It keeps telling me it cannot access the specified files ?? and when I upload a file using the paperclip in the chat it wont acknowledge them either . what is going on I am flummoxed i deleted all refreshed started a new chat tried everything but I am going round in circles I have the exact thing loaded into claude (projects) no issues at all !!
1 Like
I can’t replicate on my end. Did you try to copy paste instructions into a new one?
yeh tried everything cant get it to access any files . I can cut paste info but that defeats the purpose . as my knowledge has a load of docs to be called
I made a completely new one and got the same errors . I just asked chat gpt40_ with canvas and got the similar error like I say all was fine till i logged in this morning
chat form gpt4.0 below
Issue Summary: I experienced inconsistent behavior when interacting with file uploads in the chat system. The AI assistant provided contradictory responses regarding its ability to access uploaded files. Initially, the assistant claimed it could not access a file I had uploaded, but after repeated requests, it eventually accessed and read the file correctly. This behavior led to confusion and frustration, as it wasn’t clear why the assistant’s responses changed.
Detailed Explanation:
First Interaction:
I uploaded a file named “Bridge_.md” and asked the assistant to read and explain it.
The assistant incorrectly stated that it could not access the file.
I attempted to clarify, and after some persistent questioning, the assistant finally accessed and read the file, providing a correct summary.
Second Interaction (in a new chat iteration):
I uploaded the same “Bridge_.md” file and asked the assistant to explain it.
The assistant again responded incorrectly, stating that it could not access the uploaded file.
This inconsistency—where the assistant eventually accesses the file after initially denying it—happened in both the original and new chat interactions, only after I questioned the response multiple times.
Error Manifestation:
The assistant’s responses were not only inconsistent between iterations but also inconsistent within a single conversation.
During the same conversation, it initially denied access to the uploaded file, but after being “badgered,” it accessed it without explanation.
This led to a confusing experience where I couldn’t understand whether the assistant could genuinely access uploaded files or not.
Impact:
This inconsistency created significant confusion and made it challenging to proceed with any tasks involving file analysis.
The unpredictable responses required me to spend extra time questioning the assistant before it worked properly, which reduced trust in the reliability of the assistant.
Steps to Reproduce:
Start a conversation with the AI.
Upload a file (e.g., “Bridge_.md”).
Ask the AI to read and explain the file.
The AI may incorrectly state it cannot access the file.
Persist in questioning the AI about accessing the file.
The AI will eventually access the file and provide an explanation.
Start a new chat iteration.
Re-upload the same file.
Ask the AI to read and explain the file again.
Outcome Observed:
Inconsistent behavior both within the same conversation and between different chat iterations regarding file access.
Initially incorrect responses that were reversed only after repeated questioning.
Desired Outcome:
Consistent and reliable behavior in accessing uploaded files across all chat iterations.
If there is genuinely a restriction or limitation on file access, provide a clear and immediate message explaining the situation accurately, rather than conflicting statements.
Ensure that the AI doesn’t require repeated questioning to function properly—clarity and correct responses should happen on the first attempt.
It reads like a logic error. How do you have your instructions set to access knowledge. Can you tell it directly to read knowledge? Have you tried more than one instance? You may need to debug it is having a conflict somewhere I find numbering my instructions and examples as sub 1.1 1.2 as examples helps it understand logic flow better