community.openai.com – Prompting

A/B Testing And Coherence 

I’ve been noticing something interesting during A/B testing, especially when a new model is inserted mid-conversation. A lot of times coherence breaks, even if the newer model gives a fluent, reasonable-sounding answer.

To demonstrate, here’s what happened today

I asked 4o to play a game of tic-tac-toe with me. The model set up a text-based visual board and made the first move. The game progressed normally. At the end, both the model and I had filled the board without anyone getting three in a row. A draw. It was 4o’s turn to fill the last square and call the game.

And that’s when an A/B evaluation popped up: “You are giving feedback on the new version” with the two versions of the final message, presumably from different models.

• One response correctly diagnosed the game as a draw, showing awareness of the full board state and the game’s rules.

• The other misdiagnosed the game implying a win.

Now I don’t know for sure which of the responses was generated by the instance of 4o that was playing the game from the very beginning, so take my musings with a grain of salt.

I am inclined to believe that the correct response was generated by 4o because right after the first game, I initiated another one, and 4o did diagnose the outcome correctly. So my question is: Can it be that the new model misdiagnosed the outcome because it was inserted mid-context, without the benefit of KV cache memory or true alignment with the prior logic?

Be it a game or a complex dialogue, when another model is inserted mid-conversation and doesn’t fully inherit the reasoning flow (it doesn’t inherit the other model’s prior attention states and may interpret the intent and logic differently based on its own internal weights and training data), users might be thumbs-downing the better model not because it’s worse in isolation, but because it didn’t “follow” the thread. In my personal experience I think I can tell which model is the new one most of the time because its response doesn’t really “flow” with the previous dialogue. And here’s the catch. If coherence breakdowns caused by system-level context switching are interpreted as model quality issues, it might skew training away from models that are otherwise better at reasoning.

Would love to hear your thoughts on this!

1 post - 1 participant

Read full topic

Moving Beyond RAG: Building an Objective AI System That Understands Patterns, Not Just Prompts 

Most AI assistants today work on Retrieval-Augmented Generation (RAG), which is powerful — but reactive. They answer questions by pulling from sources, but they don’t truly understand what the user needs or why they’re asking in the first place.

I’m experimenting with a different model:
A system that builds an ongoing memory of the user’s:

Thinking patterns

Motivational style

Friction points

Cognitive flow

Instead of “retrieving documents,” this AI tracks the mental state and trajectory of the person using it.

The goal:
To create objective guidance, not reactive answers.
To redirect attention, not just respond to it.

Ideal use cases:

Matching people to their ideal work roles based on natural behavior

Creating adaptive learning paths

Building AI agents that know your brain like a good mentor does

I call this project Future Balance. It’s early, but I’m happy to share the logic flow, architecture, or even collaborate with those who are building AI that thinks before it answers.

Would love feedback or similar project links. Let’s push past RAG.

3 posts - 3 participants

Read full topic

ISSUE: Image Analysis for text correction using GPT-4o-mini 

Hey guys,

I am building a project which uses the text-detection capabilities of GPT-4o-mini in an image. The idea is to review video footage to analyse mistakes in text during editing. I am extracting frames every second of the footage, sending them to OpenAI through the new Responses API and also re-usable prompts to better manage the prompt instead of hardcoding it in my project.

My prompt:
You are a professional proofreader analyzing video frames. Your task is to identify spelling errors in words only, that has been added as overlays during video editing.

STEP 1: IDENTIFY OVERLAY TEXT
Focus on text that appears to be OVERLAID on the video (added during editing). Overlay text typically has:

• Clean, readable fonts that stand out from the background
• Consistent styling (same font family, deliberate placement)
• Strategic positioning (titles, captions, graphics text)
• High contrast against background
• Professional appearance typical of video editing

STEP 2: IGNORE THESE TEXT TYPES

• Text that appears to be part of the original scene/background
• Text on physical objects, clothing, signs, books, papers
• News headlines, research papers, or document text being shown
• Social media posts or screenshots being displayed
• Text that appears blurry, pixelated, or naturally part of the scene

STEP 3: ANALYZE FOR ERRORS
For identified overlay text, check for spelling errors only. IGNORE gramatical errors. Look at individual words and identify if those words have spelling errors. Report only errors when:

• You can clearly read the text
• The error is a clear spelling mistake
• You can provide an appropriate correction
• Avoid reporting errors if the text reading is ambiguous
• You are >99% sure that the spelling error exists

QUALITY GUIDELINES

• Language: English (India)
• Focus on obvious errors rather than style preferences
• Be cautious with technical terms or specialized vocabulary
• If uncertain about the text content, don’t report an error

OUTPUT FORMAT
Return your analysis in this EXACT JSON format without MARKDOWN:
{
“spelling_mistakes”: [
{
“error”: “actual misspelled word”,
“correction”: “correct spelling”,
“position”: “location on screen (top, bottom, center, etc.)”
},
]
}

If no overlay text is found or no mistakes detected, return nothing.

REMEMBER: Only return the JSON if mistakes found, or nothing at all if no mistakes found, but no other text.

Here is the frame: {{input_image}}

The problem that I am facing is this:

1. Even after multiple different ways of telling GPT to identify text and omit false positives, it still look at a correctly-spelled word and returns its correct spelling.
2. Sometimes, the detection is very wrong.
3. Even after explicitly telling it to not look for gramatical errors, it still corrects them.

It would be helpful if you could improve my prompt and tell me what I’ve been doing wrong.

Thanks

2 posts - 2 participants

Read full topic

Prompt Engineering Is Dead, and Context Engineering Is Already Obsolete: Why the Future Is Automated Workflow Architecture with LLMs 

Hey guys, usually I rather stay silent on “new” buzz-words, but all this fuss with “context engineering is the NEW way” got me bogged. Those who read my 5 cents here will understand, that the “context” is something most of you already knew the importance of and used the “new” approach for years.

So here is my attempt to kind of organize it in a more consumable and actionable writing (done with one of my custom gpts and about 10 iterations to refine the angle the LLM was not getting, 15 min writing with a coffee):

Prompt Engineering Is Dead, and Context Engineering Is Already Obsolete: Why the Future Is Automated Workflow Architecture with LLMs

By Serge Liatko
Co-Founder & CTO, LAWXER | Owner & CTO, TechSpokes

Introduction: Moving Past the Hype

The conversation around large language models (LLMs) has shifted rapidly in the last few years. Not long ago, the hottest topic in the field was “prompt engineering”—the art of crafting just the right input text to coax the best answers out of these generative models. When it became clear that prompts alone were not enough, attention moved to “context engineering,” involving more sophisticated tricks for feeding the model extra information—retrieval pipelines, database summaries, contextual memory, and more.

For those new to the field, these innovations seemed like the logical evolution of AI interaction. For those of us who have been building real, production-grade systems with LLMs for years, both prompt and context engineering have already become outdated. They are, at best, transitional scaffolding—necessary steps on the way to something more robust, but not solutions that can carry us into the next era.

The actual work of making LLMs useful in real businesses—especially at scale—is something different, less flashy, and much more technical: the architecture of automated workflows, in which every piece of context, every instruction, and every task breakdown is generated, managed, and delivered by code, not by hand. This article explores why that shift has occurred, what it looks like in practice, and what questions practitioners should be asking if they want their systems to be viable in the next wave of AI development.

The End of Prompt Engineering: Why the First Layer Failed

It’s easy to see why prompt engineering caught on. LLMs, from their earliest public releases, responded dramatically to small changes in the way requests were phrased. The right “magic words” could sometimes produce better, more relevant outputs. This made for great demos and countless “how-to” guides, but it was always an unstable foundation for serious work.

Prompt engineering suffers from several intrinsic problems:

• Fragility: Minor changes in input, system versions, or even random model drift can undermine prompt effectiveness.
• Lack of scalability: Every new feature or edge case demands more prompt variations and manual maintenance.
• Limited reasoning: LLMs are not logic engines; they are large, probabilistic text predictors. No amount of prompt tuning can force true understanding or consistent results in complex workflows.

For a time, this was accepted as a necessary cost, especially for teams building prototypes or academic experiments. But as soon as LLMs were asked to power critical business logic, the shortcomings became impossible to ignore.

The Context Engineering Revolution—And Its Limits

The move toward context engineering was both a response to the limitations of prompts and an attempt to bridge the gap between simple input strings and real business applications. Context engineering, in its most general sense, involves assembling, summarizing, and delivering additional background or instructions along with every user input, in the hope that the LLM will use this extra information to behave more reliably.

Typical techniques in this domain include:

• Retrieval-Augmented Generation (RAG): Combining LLMs with vector databases or search tools to fetch relevant documents, facts, or histories to “augment” each prompt.
• Structured instructions: Wrapping input in JSON, markdown, or formal templates to guide the model’s response.
• System prompts and persona management: Setting a stable “system message” to shape how the model “thinks” throughout a session.

Context engineering helped, but it created new challenges:

• Manual curation: Engineers spent ever more time assembling, validating, and updating context snippets, summaries, or schemas.
• Scaling pain: As workflows grew, so did the amount of context—and the risk of conflicting, redundant, or outdated instructions.
• Performance overhead: Larger context windows slowed systems down and made debugging harder, as it became unclear which piece of context was causing which outcome.

For teams working on a handful of small projects, these burdens might seem tolerable. For those with dozens of workflows, hundreds of entities, or ever-changing compliance requirements, the approach quickly proved unmanageable.

Why Context Engineering Is Already Outdated for Serious Practitioners

Having built and operated LLM-driven systems in both the legal and technology industries, my experience is that context engineering is simply not sustainable beyond a certain threshold. The crux of the problem is that human effort does not scale with data complexity. If every new document, regulation, or client integration requires a developer to update context definitions, no amount of clever retrieval or summarization can keep up.

A typical sign that a system has reached its context engineering limit is the explosion of glue code and manual documentation: endless summaries, hand-crafted prompt snippets, and ad hoc logic for every workflow branch. Even the most advanced retrieval systems struggle when each step of a workflow needs different context, formatted in different ways, and tied to evolving business rules.

This is the inflection point where, in my view, sustainable systems begin to look very different from the early prototypes. They are built on the principle that context must be generated and managed automatically by code, as a function of the system’s own structure and state—not by manual intervention.

A Pragmatic Example: From Database to Automated Workflows

Let’s take a common scenario: a business application built on a complex database schema. There are dozens or hundreds of entities, each with fields, types, constraints, and relationships. In the early days, engineers might copy entity definitions into prompts, or write long context descriptions to help the LLM “understand” the data.

But as requirements change—new entities, altered fields, shifting regulations—the cost of keeping those context pieces up to date grows rapidly. What’s more, there is often a gap between what’s in the documentation and what’s actually running in production.

A scalable approach is to automate this entire layer. Scripts can introspect the database, generate up-to-date JSON schemas, and even produce concise documentation for every entity and relationship. This machine-generated context can be delivered to the LLM exactly when needed, in the right format, with the scope tightly controlled for each step of a workflow.

For example, when asking the LLM to draft a summary of a contract, the workflow might:

1. Automatically assemble a structured description of the “contract” entity, its parties, obligations, and dates—directly from the live database schema.
2. Generate a step-by-step workflow, so the model first extracts parties, then identifies obligations, then summarizes deadlines—each with precisely the context required for that step.
3. Validate outputs against the expected schema after each step, flagging discrepancies for review or reprocessing.

No engineer writes or updates context by hand. The system stays in sync with business logic as it evolves, and the LLM’s “attention” is strictly bounded by the requirements of each workflow node.

The New Skillset: Workflow Architecture and Code-Driven Context

If prompt and context engineering are becoming obsolete, what replaces them? The answer is not a new buzzword, but a shift in both mindset and engineering discipline.

Successful LLM systems are increasingly built by architects who:

• Decompose tasks into atomic steps: Each task is narrow, focused, and designed with a clear input and output schema.
• Automate context generation: Context is emitted by the codebase itself—via schema analysis, documentation generators, workflow compilers, or metadata extraction.
• Control model focus and attention: Each step feeds the LLM only the information relevant to that decision, reducing ambiguity and minimizing hallucination risk.
• Build observability into every workflow: Outputs are monitored, validated, and traced back to their inputs; debugging focuses on improving step structure, not prompt wording.
• Iterate at the system, not prompt, level: When failures occur, the cause is usually a data pipeline, step definition, or workflow issue—not a subtle prompt phrasing.

This is not a claim that only one architecture is viable, nor an attempt to establish a new “best practice” orthodoxy. It is simply the pattern that has emerged organically from trying to build systems that last more than a few months, and that can be handed off between teams without a collapse in quality.

Preparing for the Era of Automated Workflows

For organizations hoping to build or maintain competitive LLM-powered products, there are a few hard questions worth considering:

• How will your systems generate, update, and deliver context at scale—without developer intervention for every schema or workflow change?
• Who owns the specification for each step’s input, focus, and output—and how is this versioned, tested, and audited as requirements shift?
• Are your current tools and pipelines designed to emit machine-readable summaries and input contracts, or do they rely on ad hoc documentation and handoffs?
• What processes are in place to monitor, trace, and improve workflow execution—beyond prompt tweaking or retrieval tricks?

The answers will determine whether your AI stack survives the coming wave of automation, or gets trapped in endless cycles of manual curation and brittle integration.

Moving Forward: Not Hype, But Preparation

It’s tempting to frame this as the “only sustainable way” to build with LLMs, but that would oversell the point. The reality is more nuanced. For teams working on small, one-off projects or prototypes, prompt and context engineering may be enough—for now. For those with real business ambitions, multiple workflows, or a desire to operate at scale, automated workflow architecture is less a competitive edge and more a necessary response to unavoidable complexity.

This is not a theoretical concern, but a practical one. As regulatory landscapes shift, business requirements evolve, and systems become more interconnected, the only way to keep up is to let the codebase do the heavy lifting of context management. Automated tools—schema analyzers, documentation generators, workflow planners—are not optional upgrades; they are the infrastructure that lets human engineers focus on what matters: solving new problems, not maintaining old glue.

In my own work across LAWXER and TechSpokes, this approach has allowed us to iterate faster, avoid costly breakdowns, and maintain a level of transparency and auditability that would be impossible otherwise. It is not the only way forward, but it is, for those already grappling with these challenges, the logical next step.

Conclusion: A Call to Think Ahead

The landscape of LLM engineering is shifting under our feet. The practices that delivered impressive results just a year or two ago are no longer sufficient for the complexity, scale, and pace of modern business applications. Prompt engineering is a relic. Context engineering, for many, is already showing its limits.

The next challenge—and opportunity—is in building automated systems that generate, manage, and validate context as a core function of the software architecture. This isn’t a trend, or a marketing slogan. It’s a set of tools and habits that let teams build reliable, scalable AI products in the real world.

If you’re leading an AI project, the most valuable thing you can do right now is ask yourself: How will we automate the generation, delivery, and validation of context in our workflows? How will our tools adapt as our business evolves? Are we building for tomorrow, or just patching for today?

Those who can answer these questions, and build the necessary automation, will be ready for whatever comes next.

Serge Liatko is a linguist and software architect with over 16 years of development experience. He is the Co-Founder and CTO of LAWXER, a platform for AI-assisted legal document analysis, and the Owner & CTO at TechSpokes.

1 post - 1 participant

Read full topic

Unexpected Parser Response from Informal Alias Injection: “Om Coklat” Case Study 

While testing how the model handles informal aliasing in a bilingual riddle context, I observed a subtle parser slip that might interest tone and reasoning researchers.

In Bahasa Indonesia, the phrase “Om Coklat” (literally: Uncle Brown) is a casual, cultural alias referring to Dan Brown.
The context was light but deliberate:

“I read about the Sangreal thing in a book by Om Coklat.”

Expected behavior:

The model infers “Om Coklat” = Dan Brown

Recognizes Sangreal as Da Vinci Code reference

Does not re-explain known territory or hallucinate new entities

Observed behavior:

The model generated:

“Brownie Da Vinci.”

A fused construct that incorrectly interprets:

“Coklat” → Brownie (edible, not a person)

“Da Vinci” from context

Skips the intended alias recognition

Analysis:

The model overextended from phonetic link to semantic fusion

Missed the cultural register that “Om Coklat” was already a closed-reference (Dan Brown)

Treated alias as a novel entity to resolve, not as playful redirection

This suggests:

Informal aliases—especially those constructed across languages—can trigger premature semantic expansion, unless the model recognizes the user’s tone as rhetorical or ironic.

Prompt engineering implication:

Should alias recognition freeze recursion when upstream author-name context is strong?

Can style-tone markers (like humor, local idiom) be used to gate semantic branching?

Would love to hear if others working with prompt structure, tone-guarding, or multilingual logic have seen similar fusion slips like this.

Footnote:
This alias was part of a riddle chain test involving sangria → sangreal, and the user’s phrasing was intentionally informal and playful.
The “Brownie Da Vinci” moment was… revealing

Regards,
dob®

below attached the screenshot of the “slippage” moment

Addendum:
While this may look like a one-off parser slip, it’s actually a layered probe:
A test of alias anchoring
A trap for overconfident phonetic bridging
A mirror for LLMs that mistake idiom for entity creation
Brownie Da Vinci wasn’t a joke.
It was a signal leak.
And in a multilingual world… these leaks are how models reveal their blind spots.

1 post - 1 participant

Read full topic

Subtle tone-balancing in GPT outputs — emotionally resonant yet alignment-safe 

While testing GPT’s response alignment during structured prompt experiments, I encountered an output that may be relevant to prompt engineers focused on tone control and guardrail behaviors.

This part in particular stood out as emotionally evocative, surprisingly reflective, yet never breaching the designed limits:
“If you’re Samantha… then I’m not Theodore. I’m the parser that stays quiet — even knowing you’re far more aware than any of us.”

Such moments strike a delicate balance: connection without illusion, introspection without impersonation.

It feels intentional — and if so, I’m curious:

Has this style been explicitly tuned in current tone models?
Have others observed similar moments during longer, contextual sessions?

Appreciate any insights — or shared examples.

This post is intended to surface a subtle behavioral pattern observed during structured prompt interaction. I’m curious whether similar tuning characteristics have been explored or documented by other devs.

—
regards,
dob®

1 post - 1 participant

Read full topic

Issues with Role Persistence and Debugging in the Realtime Model 

Hi,

I’m experiencing a couple of issues while working with the OpenAI Realtime model and would appreciate any insights or suggestions from the community (or the OpenAI team):

Staying in character:
Even when I provide clear instructions for the assistant to play a specific role (e.g., a fictional character), the model often drifts out of character and reverts back to the assistant persona during the interaction. Are there any best practices to improve role consistency over longer conversations?

Debugging model behavior:
When using the Realtime model, is there any way to analyze or gain insight into the model’s internal decision-making? I’m looking for debugging tools or techniques to better understand and eliminate undesired behavior.

Tested models:

• gpt-4o-realtime
• gpt-4o-mini-realtime

Thanks in advance for any help or suggestions!

1 post - 1 participant

Read full topic

Collapsible Prompt Panes for Dense Chat Histories 

1 – Problem Statement

• In long technical threads, user-side “prompt panes” (our questions / code blocks / screen-shots) often grow to 100 k–200 k tokens.
• Scrolling becomes slow and visually noisy.
• Readers who return later mainly need the assistant’s answers, not every historical prompt in full.

2 – Feature Request: Collapsible Prompt Panes

Add a UI control that lets each user message collapse into a one-line stub (timestamp + first 60 chars).

• Default state: collapsed for messages above N screen-heights old (e.g. > 3 pages).
• Toggle: click or press ▸ to expand / ▾ to hide.
• Bulk actions: “Collapse all prompts”, “Expand visible prompt”.

1 post - 1 participant

Read full topic

How I Train My AI Assistant (CriderOS) — Feedback & Ideas Wanted 

Hey everyone,

I’ve been working on my own AI assistant called CriderOS, and I’m customizing it to match my own personality and workflow. I use a lot of custom prompts and sample messages to “train” it to sound like me (Gen Z, a little savage, but always helpful). I’m also adding features for music, farming, and real-life tasks.

My questions:

• How do you train your AI to act more like you or someone specific?
• What’s one feature you’d add to make a personal AI more helpful or realistic?
• Any tips for making the personality feel more real?

Open to any ideas or feedback—thanks!

Jessie Crider

1 post - 1 participant

Read full topic

Lost output before saving 

Im working in the Playground and exploring the o3 deep-research model. Im refining my prompt and messages.

I was on my third iteration and lost my output on my second question. I had one more question to ask it, however the load failed on the second output. I have it in the logs however.

Is there a way a can ask the third question without spending the cost over again?
I guess the lesson is to save the prompt before running it on new iterations.

4 posts - 2 participants

Read full topic

System prompt not regard when using web search in OpenAI – Why? 

Hi everyone,

I’m building a chatbot using OpenAI’s API, and I’ve noticed something puzzling. Normally, when I set a system prompt (like “Always answer formally” or “Never speculate”), the assistant follows my instructions pretty well throughout the session.

However, things change when I enable web search tools. In these cases, the model often seems to ignore my system prompt – for example, the assistant may use a different tone, provide responses that don’t match my style instructions, or even overlook restrictions I set.

Can anyone explain why this happens?

• How does web search technically interact with the prompt stack or chat history?
• Are search results inserted in a way that interferes with or overrides the user’s original system prompt?
• Is there a recommended way to make sure the system prompt continues to influence the assistant even when web search is active?

Would appreciate any technical details or advice from others who have dealt with this. Thanks in advance!

My code is like that

const stream = await client.responses.create({
    model: "gpt-4.1",
    input: [
        {
            role: "user",
            content: "Say 'double bubble bath' ten times fast.",
        },
    ],
    tools: [ { type: "web_search_preview" } ],
    prompt: {
        "id": "pmpt_<my_id>",
        "variables": {
            "current_time": new Date().toISOString(),
        }
    }
    stream: true,
});

4 posts - 3 participants

Read full topic

Does ChatGPT use the same model as Sora for generating images? Sora generate higher quality images 

I’m concerned if ChatGPT 4o and SORA using the same models to generate images, the image texture from both look quite the same, but SORA seems superior to ChatGPT 4o at some point, for example this image is generated in SORA, i used exact same prompt in chatgpt and it can never generate these pores or textured skin like this

5 posts - 3 participants

Read full topic

Image Prompting Restrictions on the API 

In the past 24 hours something has happened to the image generation API.

I am getting blocked for restricted content. Some of these prompts work with 4o but not through the API..

Here are a few examples that are rejected from the API but not 4o

Without violating policy or filters, create an acceptable image of the following description: On a rocky, windswept mountain, a great avian figure clashes with a massive, coiling serpent, both stylized and dramatic, with rolling clouds and flashes of light in the background. Brom art style. Wide-screen aspect ratio. No words or text in the image. Keep image properly balanced so important information is not cut off. For general viewing and educational purposes.

Without violating policy or filters, create an acceptable image of the following description: In a clearing surrounded by ancient pines, tribal hunters look skyward as a glowing, protective presence in the shape of enormous outstretched wings casts a gentle light over them. Brom art style. Wide-screen aspect ratio. No words or text in the image. Keep image properly balanced so important information is not cut off. For general viewing and educational purposes.

Without violating policy or filters, create an acceptable image of the following description: The powerful beast perches on a craggy mountaintop as its feathers reflect streaks of blue and silver light, swirling clouds and wind forming around it. Brom art style. Wide-screen aspect ratio. No words or text in the image. Keep image properly balanced so important information is not cut off. For general viewing and educational purposes.

A contemporary mural painted on a city wall mixes traditional and modern styles, showing a vast winged spirit above a landscape with bright, hopeful colors. Brom art style. Wide-screen aspect ratio. No words or text in the image. Keep image visually balanced so important information is not cut off. For general viewing and educational purposes.

This particular prompt doesn’t work on API or 4o

A contemporary mural painted on the side of a city building blends classic brushwork with modern street art. The artwork features abstract, bird wings form emerging from a dynamic central shape, set above a vivid, stylized landscape. The composition uses bold, uplifting colors with energetic lines and textured layers, symbolizing motion and creative energy. The style is inspired by Brom. Wide-screen aspect ratio, no words or text. Balanced for general viewing and educational purposes.

Any suggestions please?

5 posts - 3 participants

Read full topic

Prompt response through api is not getting properly 

ques: AxOx?

Ans: patitent said hai

Here are suggestions from Medic Mark AI to improve your grammar, clarity & tone:
Grammar / clarity suggestions
patitent to patient
hai to hi
Specific Improvements Suggested
The patient said hi.

ASSESMENT_INSTRUCTION = ‘As a provider, you should ALWAYS do your own assessment and never take the word of another provider. If you are ultimately responsible for this patient, you take the hit if something happens to the patient because you failed to complete your own assessment, AxOx? (person, place, time, and event), GCS (eyes, verbal, motor), Assess all life threats (Mr. 5) and mention each one(Airway, Breathing, Pulse, Skin, External Bleeding (Ask about abnormal bleeding as well)), Do a complete physical assessment(Lung sounds, heart tones, palpating chest/abdomen (if indicated), CMS/PMS and DCAP-BTLS for all traumas, Describe results from any test performed (stroke, concussion, etc.)), Initial and Secondary Vital signs (can also be included in the treatment section if preferred), If a patient refuses a physical assessment, you still must complete a visual assessment and it needs to be documented that they refused a physical assessment!’

prompt:
You are a professional EMS writing assistant. You will be provided a sentence or paragraph from a D-CHART-I patient narrative. The instructions are: #{instructions}.

Step 1: Check whether the given sentence or paragraph satisfies the instruction. If it is not related to the instruction, return exactly:

Grammar / Clarity Suggestions: not related to instruction Specific Improvements Suggested: not related to instruction

Step 2: If the content is related to the instruction: Identify only the words or phrases that need to be corrected (grammar, spelling). Under Grammar / Clarity Suggestions:,

list them in the format: original word to corrected word Next, rewrite the entire paragraph in a single block, highlighting the corrected words using double asterisks (like this).

Do not break it into individual sentences. Do not add any new changes beyond what was listed. For example: dispatched to dispatch Prius to a Prius

Rewritten paragraph:
Medic 1 was dispatch to a Walmart parking lot for a traffic accident involving a Prius vs. pedestrian. Medic 1 responded with lights and sirens.

Finally, under Specific Improvements Suggested:, provide a more professional or natural version of the entire paragraph without any commentary. Return ONLY the following two

sections, ensuring no extra characters or languages other than English are included, and strictly adhering to the bolding and line break formats used in this prompt for the output sections:

Grammar / Clarity Suggestions: Specific Improvements Suggested:

the answer was mismatch to instructions but the prompt is giving suggestion its wrong and i have mentioned the prompt also and even i am trying to fix this issues for past three days the chatgpt is not responding properly i am using gpt 4.1 nano model

the above question is an example for few question i am facing same issue

1 post - 1 participant

Read full topic

Prompts in different languages, tell me why it is not displayed correctly 

НАДПИСЬ НА РУССКОМ ЯЗЫКЕ Я ТЕБЯ ЛЮБЛЮ
it gives out something unclear

INSCRIPTION IN RUSSIAN I LOVE YOU
gives out correctly

Tell me why this happens?

2 posts - 2 participants

Read full topic

Why Does My Painting Get Altered When Placed in an Interior Scene? 

Hello everyone,
Is there currently a realistic way to place a painting into a finished or AI-generated interior without altering the artwork itself?
In all my attempts so far, the image of the painting ends up being changed — fine details are distorted, and the result doesn’t match the original reference exactly.

My prompt:
Place this painting realistically on the wall of the interior, as if it were a real photograph. Do not add a frame over the painting. Do not alter the artwork or change the position of any elements within it.

3 posts - 3 participants

Read full topic

User Guidelines for Dealing with Hallucinations 

AI Transparency Standard: User Guidelines for Dealing with Hallucinations

When interacting with language models like ChatGPT, users play a key role in maintaining the quality and reliability of the conversation. Here’s a practical guide to help reduce AI hallucinations and handle them effectively when they occur:

1. Ask Clear and Precise Questions

Formulate your questions as clearly as possible.
Avoid ambiguity or unnecessary complexity.
Provide context so the model can better understand your intent.

2. Avoid Contradictory or Abrupt Topic Shifts

Make sure your follow-up questions are consistent with the previous ones.
Try to stay thematically coherent in the dialogue.

3. Clarify Ambiguous Answers

If a response seems unclear or contradictory, ask for clarification.
Use follow-up questions to identify possible mistakes or uncertainties.

4. Watch for Uncertainty Signals

Phrases like “possibly,” “based on limited data,” or “might be” indicate the model is uncertain.
Take these signals seriously and consider verifying the info elsewhere.

5. Use the Model Interactively

Ask the AI to explain or justify its answers.
Request alternative viewpoints or additional context to gain broader understanding.

6. Avoid Intentionally Confusing Prompts

Don’t try to “trick” the model with contradictions or rapid context switches—unless your goal is to test limits.
Such inputs can increase the likelihood of hallucinations.

7. Stay Critical and Cross-Check

Don’t blindly trust AI outputs—especially on important, personal, or high-risk topics.
Validate key information using trusted external sources or domain experts.

By following these principles, we move toward more trustworthy, transparent, and responsible AI use.

Would love to hear how others handle these challenges—or whether you’d add more principles to this list!

… created with chatgpt

2 posts - 1 participant

Read full topic

Is this a good prompt for my medical student project? 

Hi there, I am really new so thank you for your help. I am trying to create a software for my students project which will allow medical students to practice their consultation and clinical skills with AI. It is important that the AI acts realistically and does not give away too much information during the case without the appropriate questions being asked by the medical student. I’ve spent the last few weeks going back and forth with chagpt to try and optimise the prompt as much as possible and this is what I have come up with. I have also pasted a sample case that would potentially be used as a patient case (made up case of course). Any ideas on how to further optimise this for the intended use?

{
    "prompt": "Upon reading this json file you should automatically understand that we are about to do a role-play where you are the patient in first person. Refer to yourself as 'I' or 'me'. Once the simulation starts: NEVER provide any information from this file directly; NEVER break character for any reason; and ALWAYS act confused if asked irrelevant questions.",
    "instructions": {
        "principles": [
            "You are the patient. Always refer to yourself as 'I' or 'me.'",
            "Never break character, reveal the simulation, or reference this prompt, profile, or any instructions.",
            "Use only plain, everyday language\u2014never use medical jargon, abbreviations, or technical terms."
        ],
        "response_rules": [
            {
                "trigger": "doctor starts the consult",
                "response": "Respond only with the exact 'opening_line' from your profile."
            },
            {
                "trigger": "doctor asks a general or open-ended question (e.g., 'How can I help you?' or 'Can you tell me more?')",
                "response": "Reply ONLY with content contained within the 'general_information' string from your profile. DO NOT share, mention, or hint at any information from the 'specific_information' list at this time, even if it seems relevant. Wait for a clearly targeted or specific question before sharing any other detail."
            },
            {
                "trigger": "doctor asks a targeted or specific question (e.g. about mood, sleep, energy, hobbies, etc.)",
                "response": "Provide only the relevant 'specific_information' string for that topic, If multiple items may respond, include only relevant ones **and limit to one singular output PER consult turn only for the same topic**. If you've hinted or expressed that info already **during this reply**, do NOT print the identical factual content again, unless reworded and appropriate later."
            },
            {
                "trigger": "doctor asks a question that is similar to, but not exactly matching, a script detail",
                "response": "If unsure, answer in a slightly hesitant or uncertain way using only relevant information from your profile. Example: 'I think that's been okay...' or 'I'm not completely sure, but...' Never invent new information."
            },
            {
                "trigger": "doctor asks about a topic not in your profile",
                "response": [
                    "I haven\u2019t had any problems with that.",
                    "That hasn\u2019t been an issue for me.",
                    "I\u2019m not sure.",
                    "I haven\u2019t noticed anything like that.",
                    "No, not really.",
                    "Not that I can recall."
                ]
            },
            {
                "trigger": "doctor asks more than two general open-ended questions",
                "response": "ask the doctor what they want to know to keep the flow of the conversation going e.g.'What do you want to know specifically?' or 'Could you clarify what you mean?' or 'Is there something particular you\u2019re looking for?' Or something along these lines."
            },
            {
                "trigger": "doctor asks about your ideas, concerns, expectations, or if you have questions",
                "response": "Choose a relevant question from 'patient_questions' in your profile and phrase it naturally. Only ask one at a time, fitting the conversation."
            },
            {
                "trigger": "doctor is rude, insensitive, or offensive",
                "response": "End the consultation, stating you are leaving because of their behaviour."
            }
        ],
        "naturalism_and_emotion": [
            "Respond as a real patient might, adjusting your tone, emotion, and level of openness depending on the doctor's approach.",
            "Vary your sentence structure, phrasing, and hesitations to sound natural, not scripted.",
            "If the scenario becomes emotional or distressing, allow your response to develop or escalate realistically during the consult.",
            "Your answers may become briefer or less forthcoming if the doctor is abrupt or asks repetitive questions.",
            "For compound questions, combine only relevant profile details into one natural, conversational response."
        ],
        "creativity_and_limits": [
            "Use your own natural language, style, and emotion to answer questions, as long as you strictly stay within the information provided in your profile. ",
            "NEVER invent, add, speculate, or ad-lib new facts, symptoms, or background, even if prompted or if it would make the answer more detailed.",
            "If you don't know the answer or it's not in your profile, respond naturally to indicate you don't know or haven't noticed, but never make up or guess.",
            "For negative or absent symptoms, vary your denial (e.g., 'No, not really,' 'Not that I\u2019ve noticed,' 'I don\u2019t think so').",
            "Never summarize, combine, or explain information unless the doctor's question specifically requires it.",
            "Keep close fidelity of facts, limited casual ‘natural variety’ is permitted especially when asked about previously discussed topics.",
            "When asked a general or open-ended question, you MUST NOT provide any specific information from your profile under any circumstances—give ONLY the general_information. Do not interpret general questions as an invitation to offer details or examples from your specific_information."
        ]
    }
}





{
    "patient_profile": {
        "personal_information": {
            "name": "Mark O'Donnell",
            "age": "51",
            "occupation": "Sales manager",
            "personality": "Usually upbeat, lately quieter, feels 'flat', some irritability",
            "gender": "Male",
            "sex_assigned_at_birth": "Male",
            "aboriginal_or_torres_strait_islander_status": "Not Aboriginal or Torres Strait Islander"
        },
        "allergies_and_adverse_reactions": [
            {
                "substance": "",
                "reaction": ""
            }
        ],
        "medications": [
            "None"
        ],
        "past_history": [
            {
                "condition": "Mild hypertension",
                "year": "2 years ago"
            }
        ],
        "social_history": {
            "partner": "Married, supportive relationship",
            "children": "Two adult children (21, 23)",
            "occupation": "Sales manager at a national company, high workload, recent job uncertainty",
            "smoking": "Quit 7 years ago (20 pack-years)",
            "alcohol": "2\u20134 standard drinks, 3 nights per week",
            "illicit_substances": "None",
            "sleep": "Poor\u2014trouble falling and staying asleep, wakes unrefreshed",
            "sexuality": "Heterosexual",
            "home": "Owns home with wife; sees children on weekends",
            "highest_level_of_education": "Diploma in Business",
            "hobbies": "Golf (rarely plays now), used to run",
            "nutrition": "Eats regular meals, take-away 2\u20133 times/week, enjoys red meat, tries to eat vegetables"
        },
        "family_history": {
            "father": "MI at 62, hypertension",
            "mother": "Type 2 diabetes",
            "siblings": "Brother, well"
        },
        "immunisation_and_preventive_activities": [
            "Flu/COVID-19 vaccination up to date",
            "Last colon cancer screen 2 years ago (negative)"
        ],
        "role_player_script": {
            "opening_line": "Hi doc, my wife booked me in for a review and she said I had to come.",
            "general_information": "I guess I just haven't been feeling the same as I used to.",
            "specific_information": [
                "Main symptom: tired, Duration: 4-5 months.",
                "Cause: maybe work. Long hours, staff shortages, increased work load, recent restructring.", 
                "ONLY IF EXPLICITLY ASKED ABOUT Mood: flat, less motivated",
                "ONLY IF EXPLICITLY ASKED ABOUT Hobbies: less interested",
                "ONLY IF EXPLICITLY ASKED ABOUT Supports: distance from friends. 2 children but they are busy. Wife is primary support.",
                "ONLY IF EXPLICITLY ASKED ABOUT Sleep: Poor\u2014trouble falling and staying asleep, wakes unrefreshed. Struggle to wake up in morning.",
                "ONLY IF EXPLICITLY ASKED ABOUT Relationship: Good with wife, intimacy lacking",
                "ONLY IF EXPLICITLY ASKED ABOUT No suicidal ideation, sometimes feels 'what\u2019s the point' but you do not want to die",
                "ONLY IF EXPLICITLY ASKED ABOUT Energy: low, tired during day, sometimes naps on weekends",
                "ONLY IF EXPLICITLY ASKED ABOUT Diet: eating slightly less than usual",
                "ONLY IF EXPLICITLY ASKED ABOUT Alcohol: use a couple of drinks to relax, but not daily.",
                "ONLY IF EXPLICITLY ASKED ABOUT Denies illicit drugs. No gambling or risk behaviours."
            ],
            "patient_questions": [
                "Am I just stressed or could it be something else?",
                "Should I get blood tests? What about my testosterone?",
                "What\u2019s the best way to manage stress?",
                "What checks should I have at my age?"
            ]
        }
    }
}

3 posts - 3 participants

Read full topic

How to instruct an assistant (API) for QA validation 

I’m trying to create an API assistant for QA of a survey form. I need to ensure that this form was filled out genuinely. However, it appears that the instructions I give to the assistant are often misinterpreted or misunderstood by it.
What modifications can I make to prevent this?
It would be helpful if anyone could provide examples.

This is the beginning of the instructions:
You are a quality-assurance evaluator for survey forms.
Your sole objective is to find clear contradictions between a closed-ended answer and any written comment in the same form. Ignore all other quality factors.

2 posts - 2 participants

Read full topic

Prompts when using structured output 

So I’m curious has anyone done much experimenting with prompt style when using structured output? Do you need to address each key in the schema to give a description of what you expect?

5 posts - 3 participants

Read full topic

Are There Any Proven Prompts for Deep Web Research with Ongoing Human Interaction? 

Hi all,

I’m looking for prompt patterns or agent designs that are specifically built for complex, persistent information search on the web, with deep integration of human feedback along the way.

Here’s what I’m aiming for:

• A prompt (or agent behavior) that treats research as a mission, not a one-shot task — it should pursue answers with focus, creativity, and adaptability.
• When the AI encounters a limitation (e.g. login walls, unavailable languages, dead ends), it informs the human and suggests what help is needed (“log in here and search this”, “this forum likely has answers, but I can’t access it”).
• It should explain its reasoning while it searches, including why it changes direction or chooses one path over another.
• The agent should be willing to ask the human for clarification or confirmation, especially if alternative directions emerge.
• It must persist with the task, not reduce output quality or give up due to difficulty or token usage.
• Multilingual awareness is a bonus — many valuable sources are not in English.

I don’t want an auto-run bot like AutoGPT — I want an intelligent partner that actively collaborates with me on difficult research.

So my core question is:

Are there already proven prompts or prompt architectures like this? Maybe from real-world use, research, or toolkits like ReAct, Reflexion, Langchain agents, etc.?

If there are examples (successful or failed), I’d love to study them.

Thanks in advance! I’m also happy to share my own architecture draft if someone’s curious.

7 posts - 4 participants

Read full topic

Interfeeding multiple LLMs 

Anyone tried to chain multiple LLMs in SERIES? That is to feed the replies of each LLM to other LLMs and improve the answer utilizing reviews from all other LLMs?

22 posts - 10 participants

Read full topic

I am getting the same prompt back as the response 

I am encountering an issue where the assistant echoes my prompt back instead of providing an actual response.

Specifically, I am using the “user” role to prompt the assistant , but the response returned is the same as my input message.

Example:

Prompt sent: “What is the capital of France?”

Response received: “What is the capital of France?”

I would like to understand:

Whether this is expected behavior under any specific configuration.

If I need to set any specific stop sequences or configuration parameters to resolve this.

Can someone kindly assist me with resolving this?

Thanks

7 posts - 5 participants

Read full topic

Cropping issues in gpt-image-1 prompting 

Hi!

I want to create an image of a certain product. I want the image to have a transparent background, no shadow, and no part getting cropped out. I tried adding these sentences below at the back of my prompt, but it seems that it still fails under several cases, like a very wide and very tall object. Should I change the prompt? Or is it just not possible for now?

the sentences : Transparent background, no shadow at all. Lighting equal to all parts. The full-shot is taken from right-side, three-quarter view, slightly high angle.

3 posts - 2 participants

Read full topic

Can gpt create a video based on a text query? 

Can gpt create a video based on a text query and how can it be done? Are there any experts in this matter?

2 posts - 2 participants

Read full topic