![\"\"](\"https:\/\/www.campaignmastery.com\/blog\/wp-content\/uploads\/2026\/08\/kiquebg-technology-4256272.jpg\")
Image<\/a> by kiquebg<\/a> from Pixabay<\/a><\/p><\/div>\n See this post<\/a> for credits list.<\/p><\/div>\n This post has been lurking in my to-do list for so long that the ground has shifted beneath it, and not all the original advice that it was going to still holds. As of right now – July\/August 2026 – it is up to date, but that ground is unstable and the race to be better than the rest is hotly contested, so the situation can change at any point.<\/p>\n So my first tip, and the overarching one that applies to all AI – go to your preferred LLM and ask it “How has the advice and best practice for using [type of AI] changed since July 2026?” – then adjust this advice accordingly. So long as AI remains a ‘thing’ (the foreseeable future), I anticipate this becoming an annual post here on Campaign Mastery.<\/p>\n Let’s clear the elephant in the room right away – not everyone agrees with using AI in specific capacities or in general, and I understand that. To each their own. I address this question from two similar stances; when it comes to Generative (Image) AI, if it’s available on the web for a human to draw inspiration from, even if the content is copyrighted, then I think it’s fair enough for AI to be able to use it, too, to learn technique and style and gain inspiration from. When it comes to LLMs, the philosophy is similar – it’s up to you to use it ethically, and if you do so, I see nothing wrong with usage that is in line with what a human could do. If I’m designing a system mechanic for a game, I build upon past art, whether that’s copyrighted or not, but I only draw the basic principles from that past art; if I’m actively seeking to adapt some existing rules from another game into one of my campaigns, I necessarily need to understand those rules AND the rules under which my game is run, even if one or both are copyrighted.<\/p>\n Doing these things for profit instead of personal use is a whole different kettle of fish, but blaming an AI for its misuse by a human is throwing the baby out with the bathwater – at least in my opinion. But you draw your own lines.<\/p>\n One day, I hope, the AI itself will have a sense of ethics. It will have the capacity to limit the training and usage that it gets from sources it is unethical to use in certain ways, just as I can go to a clip art site and tell it’s search ‘authentic only’, or can restrict a google image search to only images available under Creative Commons Licenses. But it’s not there yet, and until it is, you and I have to plug the gap.\n<\/p><\/blockquote>\n Anything in [square brackets] is a side-note by me, including some additional advice inspired by what the AIs themselves have contributed.<\/strong><\/p>\n There are three basic ‘flavors’ of AI that are current, and each has their own nuances, capabilities and limitations. They are definitely not the same.<\/p>\n There is the granddaddy of them all (in the minds of most users), Generative AI. This takes a descriptive prompt and turns it into an image. That’s what the original version of this post was intended to cover.<\/p>\n There’s an LLM, essentially a chat-bot that analyzes and delivers text responses to text prompts, and which is capable of having a conversation. That’s what the second iteration of this post added to the first.<\/p>\n And there’s the latest cab of the rank, Agentic AI, where the AI can actually do things for you – construct software and spreadsheets and web pages and scripts. I haven’t used this at all, so there won’t be any first-hand advice to offer – but it would be negligent not to include it.<\/p>\n For the other two categories, I have a blend of hard-earned personal advice, past advice from LLMs, and current advice from a number of different LLMs – each of whom was shown the draft manuscript and asked to add to it without repeating the advice already included. So hopefully, this will be fairly comprehensive.<\/p>\n In addition to those three specific categories, I will include at the top a ‘general advice’ section for tips covering all three flavors of AI equally – but I expect this to be fairly thin as the three flavors are so different. But that’s the place to start.<\/p>\n Perplexity said that it had read the draft and had nothing to add – but then offered a list of alternative AIs to ask, which included Gemini and Claude, which had already contributed to the article – so I have to wonder just how much the free model comprehended.<\/p>\n This was actually surprising to me, because I expected it to latch onto every other ‘how to use AI’ article out there and dredge through them for the best ideas – with links to sources (which I would happily have included). But no – if it’s got nothing to add, that was explicitly defined as acceptable in the prompt.<\/p>\n However, upon reading this review, it added<\/p>\n You’re right to be underwhelmed by that response, and I appreciate the direct feedback.<\/p>\n I did read the full draft, and I interpreted your “I have nothing to add” clause quite strictly: I looked for places where I could genuinely extend or update the advice without repeating what Gemini, Claude, and ChatGPT had already covered in your text. Given how comprehensive those sections already are, my judgment was that most of what I could say would be rephrasings or minor variations.\n<\/ol>\n<\/blockquote>\n This small section has been included as a commentary on my expectations,<\/em> as much as anything else, and is not intended as criticism of the AI itself. On the contrary, I appreciated it not wanting to waste my time. And, despite claiming that it had nothing to add, it did in fact make some very useful suggestions beyond mere tips for usage, for which I have to give it double-credit.<\/p>\n Having reviewed Perplexity explicitly, fairness dictates that I at least mention my impressions of the others beyond the comments in the free list later in the article.<\/p>\n Gemini – for me, this AI hits the sweet spot between creativity, collaboration, and intellectual rigor. My go-to AI.<\/p>\n Claude – except, perhaps, when it comes to answers to be delivered in narrative \/ essay form – descriptions and the like – which is where Claude excels.<\/p>\n ChatGPT – feels like a university lecturer, who has difficulty unbending and dumbing down to an ordinary person level. Great for technical rigor. Possibly the hardest of the AIs to impress and the least likely to shower you with compliments.<\/p>\n Mistral – the differences are subtle, but it comes across as a blend of Gemini and Claude – but it contributed a number of extremely different and useful advice items. Covers the ‘blind spots’ of the other three.<\/p>\n Qwen – Helpful, and found advice that the previous contributors had all missed.<\/p>\n Meta – Quite possibly had the best big-picture view of the article as a whole. Contributed detailed advice but also structural suggestions – which is fine if you’re open to those. Far better than the impression I had of the AI from others.<\/p>\n Deepseek – Meta’s strongest rival in terms of big-picture. Again found gold that others had not suggested. Perhaps a hair more technical than Meta. They make a great 1-2 punch, and both would be elevated in the sequence if the big picture of a project was in any way uncertain or ill-defined.<\/p>\n Perplexity – I respect the ‘I have nothing to add’ – and even more the fact that it then found some worthwhile contributions to make. One of the more sensitive to nuance in prompt language, perhaps.<\/p>\n Cohere – the least impressive of the AIs consulted. See my ‘user experience’ report at the end of the article. Noticeably slower than the others, and the only one to repeatedly trigger ‘page unresponsive’ warnings.<\/p>\n Each of these is quoted sufficiently, in terms of contributions, that you can make up your own mind. The combination of all of them is definitely stronger than any one of them alone. A couple of them suggested splitting this article up, as I’ve mentioned elsewhere – this point is my second most compelling reason not to do so.<\/p>\n Several of the AIs, assessing the article as a whole, pointed out that it was, in itself, a functional demonstration of the principles it espouses – AIs as Collaborators<\/em> not Oracles<\/em> or Crystal Balls.<\/em><\/p>\n With all that out of the way, let’s get into the granular, detailed, hairy-chested specifics.<\/p>\n 1. It bears repeating: AI is a tool. Understanding its current limitations and how best to avoid tripping over them is critical to a satisfactory result. Understanding its strengths and weaknesses lets you leverage those strengths to achieve a better<\/em> result. No tool is perfect, and you need to factor that into your planning, too. Never be afraid of making your first question, “What is the best AI for [task]?”<\/em><\/strong><\/p>\n 2. Always start from the general \/ overall and then move to the specific.<\/p>\n 3. AIs don’t understand words. They translate words into conceptual ‘tokens’ which they DO understand – but that translation sometimes isn’t perfect. Some flavors are more sensitive to nuance than others; some flavors have greater token capacities than others. I’ll get into all that in dealing with each ‘flavor’. The better you understand your language, the more nuanced the results you can achieve – if something is wrong with the output, try a synonym.<\/p>\n 4. All AIs can and will make mistakes. Always have a plan in place to overcome those mistakes. Specifics will vary.<\/p>\n 5. There are studies which show that intellect and creativity decline when you use AI as a crutch. So don’t do that – use it as a stimulant for your own thinking.<\/p>\n 6. Each different AI has its own strengths and weaknesses relative to other AIs – they are NOT all the same. Find ways to leverage the differences to your advantage. One of the most effective ways to improve AI-generated advice is to ask a different AI to critique it rather than asking the same AI to rewrite it (This article is a living example).<\/p>\n From Gemini:<\/p>\n 7. Context windows are asymmetrical. An AI might be able to ingest a massive amount of data at the start of a prompt, but its output capacity (the maximum length of the response it can generate at one time) is significantly smaller. If you ask for a massive amount of generation at once, it will truncate or drop quality. Break heavy requests down into modular prompts.<\/p>\n 8. Watch out for “Sycophancy Bias.” AIs are inherently designed to be agreeable assistants. If you suggest a flawed game mechanic or a weak plot twist and ask, “Is this good?” the AI will almost always tell you it’s brilliant. To get genuine critique, format your prompt neutrally: “Identify any mechanical flaws or thematic inconsistencies in this concept.”<\/em> You can also explicitly direct the AI to “Be direct and unbiased” – this doesn’t always work, but it can help.\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 9. Give it explicit permission to disagree. By default, most AIs are tuned to be agreeable collaborators, and even telling them to “be unbiased” (as noted above) only partly overcomes this. Go further: “If you think this is a bad idea, say so plainly and explain why, before offering alternatives.” Naming the specific behavior you want, rather than just asking for honesty in the abstract, gets better results.<\/p>\n 10. Mind what you paste into free tools. Free tiers often use your conversations to train future models unless you’ve explicitly opted out in settings. Before pasting anything commercially sensitive, personally identifying, or otherwise private, check the platform’s data-usage settings – or just don’t paste it. This matters more for free tiers than paid ones, where opt-outs are usually the default.\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 11. “Prompt engineering folklore” dates quickly. Workflow refinements tend to survive model changes.<\/p>\n 12. There comes a point where another prompt isn’t improving the work, it’s just replacing one set of trade-offs with another. If the output is already good enough for your needs, stop iterating and move on. Endless refinement is one of the easiest ways to waste time with AI.\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral front end):<\/p>\n 13. Leverage model-specific strengths for niche tasks.<\/strong><\/p>\n Mistral excels at concise, direct, and technically precise outputs, making it ideal for quick fact-checking, code reviews, or distilling complex ideas into actionable insights. Use it when you need clarity over creativity – eg, debugging a mechanic, summarizing research, or validating a rule interaction. Example: “Explain the mathematical flaw in this combat balance formula in 3 bullet points.”<\/p>\n 14. Use temperature and randomness strategically.<\/strong><\/p>\n Most AI platforms (including Mistral) let you adjust “creativity” (temperature) or randomness. Lower temperature (0.2-0.5) for deterministic tasks (eg, code, rules analysis). Higher temperature (0.7-1.0) for brainstorming or narrative inspiration. Free tiers often default to medium (&Approx;0.7), which may over-embellish technical outputs.\n<\/ol>\n<\/blockquote>\n From Qwen:<\/p>\n 15. Save your prompts that work. When you discover a prompt structure that produces consistently good results – a particular phrasing, a specific sequence of instructions, a persona framing that clicks – copy it into a personal “prompt library” document. Model updates and platform changes will alter what works over time, but having a written record lets you diagnose what changed when a previously reliable prompt suddenly starts under-performing. Without that record, you’ll waste time trying to remember what you used to type. Treat it like a recipe file: the dish changes when the oven changes, but you need the original recipe to know which knob to adjust.\n<\/ol>\n<\/blockquote>\n Deepseek adds,<\/p>\n 16. Build a “failure archive” alongside your prompt library. When a prompt you expected to work produces garbage output, save both the prompt and the bad result with a one-line diagnosis (“too vague,” “ambiguous pronoun reference,” “model ignored negative constraint”). Over time, this becomes more valuable than your successful prompts – it teaches you the boundaries of each model’s comprehension, and it gives you a quick reference for what not to do when you’re in a hurry. Success tells you what works; failure tells you where the guardrails actually are.\n<\/ol>\n<\/blockquote>\n Cohere seemed to have trouble understanding “Preceding”. It conflated the ‘contribution mark’ with the following section, not as the tail of the section already underway – so instead of offering any General advice, it used this prompt as a trigger for adding to the “Generative AI (Images) Advice” section, below.<\/p>\n This is what the original form of this article was all about. It’s grown since then!<\/p>\n 1. Start with the most general description – mood \/ tone \/ style, lighting, and an overall summary of the image. Often very sensitive to nuance so use vivid and expressive terms.<\/p>\n 2. Next, the scope of the images contents – floors, walls, sky, background. Use 1-2 adjectives, 1-2 word nouns\/definitions, per item – more if you need them but try HARD not to.<\/p>\n 3. This should be followed by (same prompt) listing the elements to be present in the scene. Use 1-2 adjectives, 1-2 word nouns\/definitions, and 2-4 words to describe what the element is doing or where it is placed and which way it is facing. If you need more, you have to steal them from other categories. Elements should be listed in sequence of priority \/ importance.<\/p>\n 4. Generative AI has HUGE trouble interpreting perspective positions (“Seen from beside the stage”) and relative positions (“Left of [element x]”, “Facing element [y]”, In front of element [z], etc). Instead, locate elements by their position relative to the image (“Bottom-left corner”), and reorder the sequence as necessary. They DO generally understand “Foreground”, “Mid-ground”, “Background”, and “Close” \/ Distant”.<\/p>\n 5. As a general rule, the more you let the AI decide the content, the more usable it will be; the more specifically you describe the content, the closer it will be to your description, but the more likely it will be that something will be left out.<\/p>\n 6. As a general rule, Generative AI prompts have far fewer tokens than any other AI application. Make every one of them count.<\/p>\n 7. You might think that getting an LLM to help write your prompt would be a solution, and sometimes it is – and sometimes, it’s not. If you go this route, always edit and polish the prompt yourself. In particular, look for things that can be cut out, and\/or that can be implied.<\/p>\n 8. You will almost always think of something you’ve left out of the first version of the prompt that you use – so test it in a generator that offers unlimited images and ask yourself what’s missing or wrong before copying and pasting the prompt into your REAL first choice of generator.<\/p>\n 9. Quite often, the only way to get everything you want, exactly the way you want it, is to generate the elements separately and composite them yourself. So learn how to do that, and do it well. You can start with my series, Image Compositing for RPGs<\/a>.<\/p>\n 10. Sometimes, it’s easier to change the narrative description than it is to get the image to conform to what you had in mind going in. For example, I generated a conference room with “no doors” – and the LLM included doorWAYS leading to empty air. For multiple images. For some, I was able to clone an element in the image and resize the copy to fill the space; in others, that didn’t work, but I was able to rewrite the description so that the room ‘made sense’.<\/p>\n From Gemini:<\/p>\n 11. Negative Prompting by Exclusion: If the interface doesn’t have a specific “Negative Prompt” box, build exclusions directly into the style description. Instead of saying “a room with no clutter,” use positive terms of absence: “A sterile, minimalist, completely bare floor space.” AIs struggle with the word “no” or “not” because the token for the item still appears in their conceptual map, making them likely to draw it anyway.<\/p>\n 12. Leverage Seeds for Iteration: If you find an image that is 90% perfect but a character has six fingers or the lighting is slightly off, look for the “Seed Number” in the image metadata (if the platform exposes it). Re-prompting with the exact same seed number and a minor text tweak tells the AI to use the exact same mathematical starting noise, ensuring you don’t lose the core layout of the image you liked.\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 13. Build a reference set for recurring subjects. If a character or location needs to reappear across many images, don’t just repeat the text description – feed the AI your best accepted image back in as a reference alongside the new prompt (“keep this character’s face and armor, change the pose to…”). Text alone drifts over multiple generations; a reference image anchors it.<\/p>\n 14. State the aspect ratio and framing every time, explicitly. Don’t rely on the platform’s default. “16:9, wide establishing shot” versus “square, close portrait” changes what the AI chooses to include or crop, sometimes substantially – this is a separate lever from the content description itself. [I personally like to actually state the pixel width and height, explicitly; the AI doesn’t always listen].\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 15. Decide what matters before you prompt.<\/p>\n Not every image needs every detail to be correct. Before writing the prompt, decide which two or three features absolutely must survive generation. Those become the emphasis of the prompt. Everything else is negotiable. Trying to make every aspect equally important usually means nothing receives sufficient emphasis.<\/p>\n 16. Don’t fight a model’s strengths.<\/p>\n Different image models have noticeably different visual biases. If one consistently produces painterly images while another excels at photo-realism or fantasy illustration, work with those strengths instead of spending twenty prompt revisions trying to force one model into behaving like another.\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral):<\/p>\n 17. Use “style reference” images for consistency.<\/strong><\/p>\n If a platform supports it (eg, Leonardo, MidJourney via Discord), upload a reference image and weight its influence (eg, “–sref [URL] –sw 0.7” in some tools). This locks in art styles, color palettes, or character designs across multiple generations better than text alone.<\/p>\n 18. Exploit “inpainting” for iterative fixes.<\/strong><\/p>\n Instead of regenerating an entire image to fix a small flaw (eg, a misplaced object, extra finger), use inpainting tools (available in Freepic, Leonardo, and others) to:<\/p>\n 1. Mask the problematic area. Pro tip: Some tools let you invert the mask to protect the rest of the image from drift.<\/p>\n 19. Test “negative prompting” with synonyms.<\/strong><\/p>\n If the AI ignores “no X”, try:<\/p>\n ▪ “X-free” (eg, “text-free”, “watermark-free”). Qwen added:<\/p>\n 20. Prompt word-order effects vary by platform, and this is not a bug you can fix. The same six words rearranged will produce a different image on Nanobanana 2 versus Freepic versus Leonardo, because each model’s attention mechanism weights token position differently. If you’ve written a prompt that works beautifully on one platform and then port it to another, expect to re-tune the order of your descriptors, not just the words. Keep a brief per-platform note: “Platform X wants style first, subject second; Platform Y wants subject first, lighting last.” You’ll save yourself a dozen wasted generations every time you switch tools.\n<\/ol>\n<\/blockquote>\n Meta AI contributed:<\/p>\n 21.Use an LLM as your art director, not just your prompt writer. Before you generate, paste your prompt into an LLM and say: “You are an art director for tabletop RPGs. What will this prompt likely get wrong, crop out, or over-emphasize? Rewrite it to front-load the 3 things that must survive.”<\/em> That pre-flight check catches token-waste that you won’t see until you’ve burned 5 generations. [Tip: Tell the LLM what Generative AI the prompt is to be written for – refer 20, above!]<\/p>\n 22. Name the light source, not just the lighting. “Cinematic lighting” is vague to every model. “Single torch on stone wall to camera left, warm flicker” gives the model a physical anchor for shadows and falloff. This alone fixes 80% of “why is the dungeon evenly lit like an office” problems.<\/p>\n 23. Keep a ‘rejected but useful’ folder. That character portrait with the wrong armor but perfect face? Don’t delete it. That’s your style reference next week when point 13 from Claude fails because you have no good anchor image. Label it with what did work.\n<\/ol>\n<\/blockquote>\n From Deepseek:<\/p>\n 24. Name your unwanted style explicitly, not just the subject. “Photorealistic” is a positive request; the model will happily give you photorealistic everything. If you want a painterly illustration, don’t just ask for “painterly” – also specify “not photorealistic, not 3D render, not CGI.” Explicit exclusions in the style layer (rather than the subject layer) are more reliably respected than “no X” attached to objects. The model’s style bias is a separate attention channel from its subject attention; treat them as two distinct levers.<\/p>\n 25. When generating groups, specify the number of distinct figures and their relationship to each other in the first sentence, not buried in the middle. If you say “three adventurers – a dwarf, an elf, and a human – standing in a tavern” at the start, you’ll usually get three figures. If the group composition appears after the background, lighting, and mood descriptions, the model often decides “three figures is a suggestion” and gives you two, or four, or a crowd. Front-load the non-negotiable count and identities; everything else is negotiable after that.\n<\/ol>\n<\/blockquote>\n From Cohere Playground:<\/p>\n 26. Keep a “prompt library” of successful prompts per platform. When you find a prompt that reliably produces the desired style, save it with notes on platform, model, and any special parameters (eg, “–ar 16:9 –s 750”). Reusing known prompts reduces trial-and-error.<\/p>\n 27. Use explicit composition cues like “rule of thirds” or “golden spiral” if the model respects them. For example: “Place the main character at the intersection of the left third and upper third.” This guides placement more reliably than vague positional terms.<\/p>\n 28. Specify a color palette with hex codes or descriptive terms. Example: “Dominant colors: #2E5A3A (deep forest green), #F4E8C1 (warm parchment), accent #C73E1D (crimson).” Many models will honor a limited palette better than a generic “green and brown.”<\/p>\n 29. When you need a series of related images (eg, a map, a character portrait, and a location shot), generate a base layout image first, then feed that as a reference image for subsequent generations. This anchors style and composition across the set.<\/p>\n 30. If a free tier limits resolution or adds watermarks, generate at the highest allowed resolution, then upscale using a separate tool (eg, Real-ESRGAN, waifu2x) or crop to remove watermarks. This gives you higher-quality final assets without paying for premium tiers.<\/p>\n 31. When testing multiple variations, generate them in batches using a consistent seed range (eg, seed 1000-1010) and then compare side-by-side. This makes it easier to spot subtle differences and pick the best version without re-prompting from scratch each time.\n<\/ol>\n<\/blockquote>\n From Meta:<\/p>\n 1. The stubborn image error:<\/strong> You’ve reworded 6 times and the same unwanted element persists – extra limb, wrong era gun, text. That’s not a prompting failure, that’s the model’s bias winning. Stop prompting. Switch to masking\/inpainting for that one element, or generate that element on a plain background [transparent makes it easier] and composite. Per point 9, compositing is not failure, it’s the intended workflow at this level of specificity. [You may have to use a different AI for the inpainting.]<\/p>\n 2. The drifting set:<\/strong> Image 1 was perfect, Image 4 has lost the armor color, Image 7 has lost the face. That’s your cue that you’re relying on text alone for continuity. Re-anchor: feed Image 1 back in as a style\/reference image at 0.6-0.8 weight and reduce your text prompt to only what changes.<\/p>\n In both cases, your time check is 8 generations. If you haven’t moved forward in 8 tries, change technique, not words.\n<\/ol>\n<\/blockquote>\n Cohere added:<\/p>\n 3. Maintain a “generation log” that records prompt, seed, model version, and outcome. When drift appears, you can backtrack to the last known good generation and re-anchor from there, rather than starting from scratch.<\/p>\n 4. If a particular element (eg, hands, faces) repeatedly fails, generate that element in isolation using a model known for anatomical accuracy (eg, MidJourney with `–ar 1:1` and `–style raw` [but note Midjourney is not free]). Then composite it into the main image.<\/p>\n 5. Use prompt weighting or token emphasis (eg, in Midjourney use `((…))`, in some tools you have to use `**…**`) to force the model to focus on critical details and suppress problematic ones. This can reduce unwanted artifacts without sacrificing the overall composition.<\/p>\n 6. When you exceed the effective token budget for negative prompting, split the workflow: first generate a base image with minimal exclusions, then apply inpainting to remove or replace the offending elements. This often yields cleaner results than trying to list every exclusion in one prompt.<\/p>\n 7. Set a hard limit of 8 generations per iteration (as suggested in point 2). If you haven’t resolved the issue by then, switch to a different model or technique (eg, from a text-only generator to a compositing workflow). This prevents sunk cost and keeps momentum.\n<\/ol>\n<\/blockquote>\n From Meta AI:<\/p>\n 1. For RPG use, the real skill ceiling isn’t prompting, it’s knowing when to quit. If you’ve regenerated more than 8 times trying to get two specific things to coexist in one frame (eg, “dragon AND tavern interior AND correct scale”), you are now in compositing territory. Generate the dragon separately on black, generate the tavern, and put them together yourself. Your time is the token you should be counting. [Unless earlier attempts are incredibly close, I probably wouldn’t last through 8 attempts – 4-6 is my limit].<\/p>\n [2. Get generative AIs to make specific element images on a transparent background. “Four isolated medieval-fantasy chairs facing away from the viewer, transparent background, with shadow from an overhead source further away that the chairs” lets you pick from the best and use it.]<\/p>\n [3. Using one AI to generate all the image elements permits felicity of visual style. Sometimes, though, one simply can’t seem to understand what you want (or stubbornly refuses to create it). Use multiple generative AIs as necessary, and use the elements provided by one as a visual reference to the next.]\n<\/ol>\n<\/blockquote>\n From Cohere:<\/p>\n 4. Maintain a “composite library” of pre-generated transparent elements (characters, props, backgrounds). Tag each with style, lighting, and scale notes. Reuse these assets both within and (where appropriate) across campaigns to reduce generation time.<\/p>\n 5. When integrating multiple elements, align their lighting by generating them under a shared “light source” description (eg, “late afternoon sun from the upper left”) or by manually painting consistent shadows using a Multiply layer at low opacity.<\/p>\n 6. Use a “color match” pass after compositing: apply Photoshop’s Match Color or a similar tool to ensure the overall hue and saturation are harmonious across disparate generated pieces. [Search google for how-to with your preferred editing tool. Be sure to specify the version of the software, because one version can vary wildly from another at this level of detail.]<\/p>\n 7. For complex scenes, create a quick storyboard of individual elements first (background, mid-ground, foreground). Composite in depth order, using layer masks to preserve edges and avoid overlap conflicts.<\/p>\n 8. If a generated element includes an unwanted watermark, generate a duplicate with a transparent background or use inpainting to erase the watermark before compositing.<\/p>\n 9. Apply a subtle texture overlay (parchment, stone, fabric) to unify visually distinct elements and give the final image a cohesive “world” feel.<\/p>\n 10. Keep a simple text log of your compositing steps (eg, “Background: tavern interior, seed 123; Dragon: isolated, seed 456; Lighting: added soft rim light”). This log speeds up recreating or tweaking the scene later.\n<\/ol>\n<\/blockquote>\n LLMs are the hot ticket right now in terms of everyday usage. They are great for brainstorming, research, and finding specific solutions to problems. They are okay for game mechanic creation. They are appalling at basic arithmetic, and worse at higher math (there are hopes that this will change sooner rather than later, but in general, Agentic AI is better for these things).<\/p>\n 0. I’m tagging this as “tip zero” because it is that fundamental, even though it’s not about AI interaction directly – some browsers load the most recent prompts and responses into memory to make switching between them faster. This can cause timeouts and overloads when you have a lot of them. As soon as you are SURE that you have finished with a thread, delete it.<\/p>\n From Gemini:<\/p>\n 1. Assign a Persona Explicitly: Do not just ask for information. Establish the AI’s role in the very first sentence. Tell it: “You are an expert tabletop RPG systems designer specializing in balanced combat mechanics,” or “You are a ruthless editor critiquing a sci-fi narrative.” This forces the LLM to pull from a more targeted subset of its training data, yielding sharper, less generic responses.<\/p>\n 2. Use “Chain of Thought” Prompting for Mechanics: If you are asking an LLM to build a game mechanic or resolve a complex narrative timeline, explicitly tell it to “Think step-by-step out loud before providing the final answer.”<\/em> Forcing the AI to generate its reasoning path sequentially reduces logic errors and math hallucinations drastically.<\/p>\n 3. The “Sandwich” Formatting Rule: When pasting a large chunk of text to be analyzed or edited, wrap the source text in clear delimiters (like [START TEXT] and [END TEXT]). Put your instructions both at the very top and repeated briefly at the very bottom. LLMs suffer from “loss in the middle,” meaning they pay the highest attention to the tokens at the absolute beginning and the absolute end of the prompt window.\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 4. Use the platform’s persistent custom instructions, if it has them, rather than re-establishing persona and preferences every session. Most major LLMs now let you set standing instructions (“always answer in metric,” “you’re helping with a tabletop RPG campaign, assume fantasy genre knowledge”) once, at the account level, so you’re not repeating yourself at the top of every new thread. Check your settings menu – it’s not the same feature as Gemini’s per-prompt persona assignment above, and the two work well together.<\/p>\n 5. Ask it to restate the task back to you before it starts, on anything non-trivial: “Before you begin, summarize what you understand I’m asking for.”<\/em> This catches a misread instruction while it’s cheap to fix, rather than after it’s produced 2,000 words down the wrong path.\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 6. Separate thinking from formatting.<\/p>\n When asking for something substantial, first ask the model to solve the problem. Then, in a second prompt, ask it to present the answer in the desired format. Separating analysis from presentation often produces better results than asking it to perform both tasks simultaneously.<\/p>\n 7. Tell it what success looks like.<\/p>\n Instead of simply describing the task, describe the finished product. “I’ll consider this successful if…” gives the model a concrete target against which it can evaluate its own draft.\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral):<\/p>\n 8. Preempt hallucinations with “citation mode” or source requests.<\/strong><\/p>\n If the platform supports it (eg, Perplexity, Mistral’s Le Chat with web search), force the AI to cite sources for factual claims. For Mistral without tools: “List all assumptions you’re making in this answer” to expose gaps.<\/p>\n [I have also found it useful at times to include the initial instruction: “Validate all assumptions with me before incorporating them into your thinking.”]<\/p>\n 9. Use “few-shot prompting” for custom formats.<\/strong><\/p>\n If you need outputs in a specific structure (eg, stat blocks, tables), provide 2-3 examples of the desired format before your actual request. Mistral adapts quickly to patterns.<\/p>\n Example: 10. Request “contrarian” or “devil’s advocate” modes.<\/p>\n To break sycophancy bias, try: “Act as a critic who dislikes this idea. Give me 3 harsh but fair objections.”<\/em> or “You are a player who hates this mechanic. How would you exploit or break it?”<\/em>\n<\/ol>\n<\/blockquote>\n Qwen suggests also:<\/p>\n 11. Specify the output shape before the content. Before your substantive request, tell the AI the physical form of the answer you want: “Respond in a three-column table,” “Give me exactly seven bullet points, no more,” “Write this as a 200-word paragraph followed by a numbered list.” Constraining the container changes how the model organizes its thinking, not just how it looks on screen. An LLM asked to “compare X and Y” will ramble; the same LLM told “compare X and Y in a four-row table with columns for Feature, X, Y, and Verdict” will produce something you can actually use at the table without reformatting. This is distinct from ChatGPT’s point 6 (separate thinking from formatting) – that’s a two-step process; this is a one-step structural constraint that improves the thinking itself.\n<\/ol>\n<\/blockquote>\n [You can suggest am approximate word limit to answers from an AI, and some of them WILL adhere to it. Other’s won’t. I generally don’t do this, because I always wonder what important stuff had to be left out to meet the word limit.]<\/p>\n Meta AI weighed in with:<\/p>\n 12. Treat the model as a collaboration partner with asymmetric memory. It will remember everything in this chat perfectly, but nothing from your other chats unless you tell it. When you switch models, you are switching collaborators who have never met. The fastest way to waste time is to assume continuity that isn’t there.<\/strong> Start each new thread with a 2-sentence reset: who you are in this context, and what “done” looks like.<\/p>\n 13. The second answer is almost always better than the first, but the fifth is rarely better than the second. The best approach is iterative compression: first answer = breadth, second answer = you pointing at what you actually liked, third answer = tightening. After that you’re usually just trading one compromise for another (as ChatGPT noted in General Advice point 12). Learn to stop at “good enough for this job”.<\/p>\n 14. Free tiers are now throttled on reasoning, not just length. Since mid-2025 most free LLMs will give you a fast, shallow answer unless you explicitly say “take your time and think this through”<\/em> or “use extended reasoning”<\/em>. If an answer feels glib, ask for the slow version before you decide the model is dumb.<\/p>\n 15. Ask for the failure mode first. Instead of “Create a balanced encounter system”, try “What are the 3 most common ways encounter balance systems break in long campaigns? Now design to avoid those.” It forces the model to retrieve problems before solutions, which dramatically reduces sycophancy and generic advice.<\/p>\n 16. Use the model to interview you. For complex design tasks, prompt: “Ask me 5 clarifying questions that would let you give a 10x better answer to this question ” This puts the burden of discovering unknowns on the model, not you. You often don’t know what you haven’t specified.\n<\/ol>\n<\/blockquote>\n Deepseek offers:<\/p>\n 17. Use the “five-minute rule” for time-sensitive queries. If you’re asking about something that might have changed recently (pricing, model capabilities, platform features, errata), add: “If your training data cuts off before [date], say so and I’ll check manually.” Models don’t know their own knowledge cutoff unless prompted to consider it, and they’ll confidently answer with stale information rather than admitting uncertainty. This is especially important for the article’s own “ask how advice has changed” tip – if you don’t force the model to acknowledge its cutoff, it will simply invent an answer.<\/p>\n 18. When the AI gives you a numbered list, ask it to re-order by priority, not by category. Models default to thematic grouping (“here are three technical issues, here are three narrative issues”) which can obscure which problems matter most. A follow-up prompt – “Now rank those items from highest-impact to lowest-impact and explain the ranking” – forces the model to do value-weighting, which is a different cognitive process from generating the list in the first place. You’ll often discover that the item buried at #7 is actually the one that breaks everything. [Don’t ask it to do both jobs at once, it’s better to let it focus.]\n<\/ol>\n<\/blockquote>\n Cohere points out,<\/p>\n 19. Much of the advice under “Continue Chat Threads” (below) can also apply to using LLMs in general. In particular, entries 12, 15, and 16.\n<\/ol>\n<\/blockquote>\n 1 .A master briefing document offers huge advantages – not least of which being the ability to completely reset the conversation; just delete any old chats on the subject and re-upload the document. LLMs have an upload capacity capable of accepting about 30,000 words (100,000 tokens) at a time, but I limit briefing documents to 10,000 or preferably less; this leaves 10,000 for the LLM to analyze the document without trimming parts off that it thinks it is finished with, and 10,000 for it to think about the contents in relation to the objective of the chat session. If one ever reports that the text appears to break off, mid-sentence, that’s a sure sign that you’ve exceeded this restriction.<\/p>\n 2. At one point, the most efficient approach was to copy and paste text into a prompt, 10-15,000 words at a time. When I did so with the Dr Who campaign’s master briefing document at one point, following the 5th such upload, it was already hallucinating content and losing track of material that had been explicitly stated. This is when the capacity for accepting correction becomes CRITICAL. This problem can be reduced by delivering uploads a few at a time – if the LLM lets you upload 9 documents in a single prompt, upload no more than 3-4 (half, round down).<\/p>\n 3. Briefing documents should start with a general overview, progress to the AI’s role and any general restrictions, and then talk about specifics. This is less important than it used to be, but be methodical and structured and logical. DON’T start high, drill down to specifics of one item, then bounce back up to the big picture and start on the next<\/em> set of specifics.<\/p>\n 4. ALWAYS include the instruction to “Ask questions if anything needs clarification, don’t make assumptions on your own.” See also LLM Advice 16.<\/p>\n 5. DO use some sort of structural prompt. There are two formats that I use – for the Doctor Who Master Briefing, I’ve just numbered every paragraph. For the Zenith-3 Master Briefing, I have sections “A”, “B”, “C”, and so on, each with a heading, and with numbered paragraphs in each “A1”, “A2”, “A3”, etc. This helps the LLM build a road-map of how the content relates to each other.<\/p>\n 6. Every clarification needed or error that you notice is a sign that part of your master briefing can be improved.<\/p>\n 7. After providing any necessary background material needed, get specific with a plan of action for what is actually required of the AI. If there’s more that you need to explain, you can do so as you go.<\/p>\n From Gemini:<\/p>\n 8. Create a Glossary Token Dictionary: If your campaign uses unique sci-fi or fantasy proper nouns (eg, specific alien races, fictional minerals, or unique magic systems), include a dedicated “Glossary” section at the top of the document. Explicitly define what these words mean in relation to real-world equivalents (eg, “Xylok: A sentient crystalline mineral, behaves like a organic semiconductor”). This stops the LLM from trying to translate your fictional worldbuilding into generic real-world dictionary definitions.\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 9. Keep dated versions of your briefing document, don’t just overwrite the old one. When you revise it (per point 6, above), save it as a new file rather than editing in place. This gives you a paper trail if a later version turns out worse in some way you didn’t anticipate, and lets you diff what changed when something starts going wrong.\n<\/ol>\n<\/blockquote>\n \nBonus hint: there is freeware out there that compares two text files – it can be priceless for this kind of task. The one I use is WinMerge.\n<\/p><\/blockquote>\n 10. Distinguish facts from preferences.<\/p>\n If your briefing document mixes immutable facts with stylistic preferences, label them separately. Models generally do better when they know which instructions are absolute requirements (“Character X cannot know this information”) and which are merely preferred (“Use Australian spelling where practical”).\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral):<\/p>\n 11. Use hierarchical headers for navigation.<\/strong><\/p>\n Mistral (and most LLMs) parse Markdown headers (##, ###) to infer document structure. In briefings:<\/p>\n ▪ ## for major sections (eg, “World Lore”). This helps the AI summarize or locate info without token waste.<\/p>\n 12. Embed “meta-instructions” in the briefing.<\/p>\n Add a section like:<\/p>\n ▪ ### AI Guidelines This reduces repetitive corrections. [Every correction effectively uses two tokens to do the work of one – the original and the correction. If you have to make multiple corrections, this can add up fast.]\n<\/ol>\n<\/blockquote>\n Qwen also advises,<\/p>\n 13. Include a “Decisions Already Made” register. In a long-running campaign briefing, there will be questions you settled three sessions ago that the AI will be tempted to re-litigate or re-suggest alternatives for. Add a short, explicit section: “The following are settled and not open for revision: [list]. Do not propose changes to these unless I specifically ask.” This is different from Claude’s point 10 (distinguishing facts from preferences) – those are constraints; this is a closed-doors list. Without it, you’ll spend tokens and patience swatting down suggestions for things you decided weeks ago, and each re-litigation subtly erodes the AI’s confidence that your other stated decisions are actually firm.\n<\/ol>\n<\/blockquote>\n Meta suggests,<\/p>\n 14. Add a “How to talk to me about this document” section at the end. Literally tell the model: “When I say ‘lore’ I mean Section B. When I say ‘current party’ I mean Section D2. If I ask for ‘canon,’ quote verbatim, if I ask for ‘ideas’ you may invent.” This gives you a command language that survives across reboots and stops the model from treating every reference as a search problem.\n<\/ol>\n<\/blockquote>\n [15. Ask an LLM to review your briefing document. You want any suggestions it has regarding logical sequencing and missing or contradictory content – the latter points to something that needs clarification in the document. Remember, this is aimed at an LLM, not a human reader; some suggestions will almost ALWAYS make it better for the first at the expense of the second.]<\/p>\n 16. Put your anti-lore at the bottom, not the top. Models pay more attention to beginnings, so don’t waste that prime real estate on “Things that are NOT in this world”. Put prohibitions and “never do this” in a final Appendix titled “Hard No’s”. The model will still respect them, but you won’t have front-loaded it [burning precious tokens] with what you don’t<\/em> want.\n<\/ol>\n<\/blockquote>\n Deepseek had nothing to add to this section.<\/p>\n Cohere added a number of points – but they were all repetitions of either the general advice or advice offered in other sections.<\/p>\n 1. Lots of people go to an LLM and simply start typing. This is possibly the worst possible approach. Instead, look for a list of past conversations and continue a prior chat session if it is relevant to the task at hand.<\/em><\/p>\n 2. When you are SURE that you no longer need a chat session, look for a way to delete it. To facilitate quick response, browsers can cache the first page or two of ALL past chat sessions, and that can lead to time-outs and ‘page unresponsive’ errors.<\/p>\n 3. Social chitchat is all well and good but it eats up tokens while contributing nothing to the project. Be respectful but firm; you are talking to an assistant \/ professional collaborator, not to a best friend.<\/p>\n 4. LLMs are lousy at perceiving the passage of time between sessions. Every prompt and response is time-stamped, so they shouldn’t make this mistake, but every prompt is “Now” to them unless you say otherwise. If you can,<\/em> (sometimes you can’t), treat this as a harmless delusion on their part. The main time when this is not a viable option is when you have a DEADLINE to meet<\/em><\/strong> or when the time interval provides necessary context. In which case, be explicit up-front.<\/p>\n 5. At the top of every session in an ongoing chat, repeat the instruction that the LLM should “focus attention specifically on this conversation thread only.”<\/p>\n 6. Most chat sessions are given an abbreviated title by the LLM. Sometimes these are very good, sometimes they are abysmal – but they can usually be edited or renamed. Use this ability to ensure that YOU can glance at the title and know what the thread is about.<\/p>\n 7. There are limits to how far back in an ongoing conversation you can scroll back. No chat session is complete until you have archived on YOUR hard disk anything that you want to keep. Copy and paste prompts and responses into a working document on your computer.<\/p>\n 8. When a conversation gets REALLY long, this technique can fail – attempts to copy a response produce a “browser not responding” error. When this happens:<\/p>\n 9. Eventually, in a long conversation, an AI may start to hallucinate content that was never there, or lose track of specifics that were. This happens when the cognitive load required to keep track of the thread exceeds the token limit allocated to the conversation, causing the LLM to summarize the older content – from 100,000 tokens down to, say, 10,000 tokens. Some LLMs are better at picking up where they left off after integrating corrections into this summary – Gemini is better at this than most. So correct it as necessary and you may be able to carry on.<\/p>\n 10. When the summary approaches the token limit, things can get really broken. You correct one hallucination only to trigger another. When this happens, you need to try starting a rebooted thread as described in 8, above.<\/p>\n 11. And when that’s not enough, you need to update your master briefing document with your own summary of what’s important, delete the old thread, and start over from that position in the process.<\/p>\n 12.We alll have random queries from time to time, not part of any ongoing thread. Because these can accumulate to the point where the clog up your browser (as described in (8) above), I also put these into a dedicated continuing chat session, “Miscellenious Questions”.<\/p>\n 13. Because there are limits to how far you can scroll back, when something comes up that I want to save (perhaps for an article here at Campaign Mastery), I will start a new discussion thread, “Miscellenious Questions 2” or “-3” ot whatever, rather than risk that happening. When whatever caused this behavior is dealt with, I can safely delete the old “Miscellenious Questions” thread.<\/p>\n 14. It’s also worth taking the time to ‘clean house’ of old conversations occasionally. Newer concepts in AI permit infomation from one discussion thread to bleed into another – so, once you don’t need it any more, get rid of it. I also make it a practice, where it’s important, to state in my first prompt of a session, “Confine your attention to the contents of the thread named [X]” – it’s not perfect, but it does help with this problem.<\/p>\n From Gemini:<\/p>\n 15. The “Temperature Reset” Prompt: If a thread has gone on for a while and the AI is starting to sound repetitive, robotic, or overly fixated on a specific phrase it coined five turns ago, use a behavioral circuit breaker. Prompt it: “Let’s clear our stylistic slate. For the next response, adopt a completely fresh prose style, eliminate recurring catchphrases from this thread, and approach the problem from an entirely new angle.”<\/em>\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 16. Consider running two parallel threads for the same project: one for drafting, one for critique. A thread that’s been enthusiastically helping you build something all session tends to stay enthusiastic about it, even when you ask it to find flaws – it’s anchored on the collaborative momentum. A fresh thread, given only the finished output and asked to critique it cold, is more likely to actually find the problems. [Or you could use a completely different AI, ensuring that there is no carryover].\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 17. Periodically ask the AI to summarize the current state of the project.<\/p>\n Every so often, ask it to summarize the project’s assumptions, decisions, unresolved questions, and current objectives. Besides creating a useful checkpoint for yourself, this often exposes misunderstandings before they become embedded in later work.\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral)<\/p>\n 18. Use “thread anchors” for context recovery.<\/strong><\/p>\n When rebooting a long thread, start with: “This is a continuation of [Thread Name]. Key open questions: [1], [2]. Current status: [X]. Next step: [Y].”<\/em><\/p>\n AI context windows (32K-128K tokens, depending on model) handle this well, with explicit anchors preventing drift.<\/p>\n 19. Export threads as JSON for post-processing.<\/strong><\/p>\n If a thread gets too long, some platforms (or browser extensions) let you export chats as JSON. Use this to:<\/p>\n ▪ Search\/analyze the history locally. Qwen adds,<\/p>\n 20. Recognize the “compression loss” threshold. Points 9-11 describe what happens when a thread gets too long, but they treat it as a binary: things work, then they don’t. In practice, there’s a gradual degradation [that you may not notice unless you are looking for it]. Each time the LLM silently compresses older material to stay within its context window, it sheds specificity first<\/em> and facts second.<\/em><\/p>\n You’ll notice it as the AI starting to paraphrase your established terminology, conflating two characters who share a faction, or referring to events in vaguer terms than you originally stated. When you catch that softening of precision – even if no outright hallucination has occurred – that’s your early-warning signal to archive and reboot, not to push on for “just one more prompt.” By the time you get actual hallucinations (point 9), you’ve already been operating in degraded mode for several exchanges.\n<\/ol>\n<\/blockquote>\n Meta suggests,<\/p>\n 21. Another indicator of approaching the “compression loss” threshold: Watch for pronoun drift. The earliest symptom of compression isn’t hallucination, it’s the model switching from “Jorah” to “the captain” to “he” to “the former soldier”. When specific proper nouns start dissolving into generic roles, you’re already losing specificity. That’s your cue to reboot, per above.<\/p>\n 22. Don’t extend – summarize and reboot with intent. The article already notes degradation. My practical rule: after &Approx;15 substantive exchanges, ask the model “Summarize this conversation as a briefing document for a new instance of yourself to continue the work. Include: decisions made, open questions, and current objective. Keep it under 600 words.” Then start a new chat with that summary pasted. You lose flavor, you keep architecture.<\/p>\n [While that’s good, procedurally, when it’s time to reboot the conversation, I don’t agree with it wholeheartedly – especially when working on BIG projects. And, sometimes, 600 words isn’t enough, there’s too much left out – if that’s the case, tell the LLM “That’s too compressed. Make it 1000 words and try again.” And ALWAYS manually edit the prompt it gives you, because summarizing and compression are prime territories for hallucinations to occur, and you DON’T want to embed those in your restarting point!]<\/p>\n 23. Label your reboots. When you do reboot, start with “This is a continuation of [Campaign Name] Thread 3, picking up after we finalized X. Previous summary follows.” That label becomes an anchor the model will reuse when it gets confused, which is more reliable than relying on token memory.\n<\/ol>\n<\/blockquote>\n Deepseek recommends also:<\/p>\n 24.When rebooting, paste the last three exchanges verbatim, not [just] the summary. The summary tells the model where you intended to be; the last three exchanges tell it the actual state of the conversation, including any drift or misunderstanding that crept in before you noticed. Summaries preserve intent; raw history preserves reality. Use both, but if you have to choose one, use the raw history. (The summary is for you, not for the new instance.)<\/p>\n 25. Rename threads by their unresolved question, not their topic. A thread titled “Magic Item Balance” tells you the subject; a thread titled “Does the 10-minute casting time break the economy?” tells you why you’re still in that thread. When you return to it after a gap, the unresolved-question title gives you immediate context for what’s actually at stake, saving you a re-read of the last several exchanges.\n<\/ol>\n<\/blockquote>\n Cohere added a number of points – but they were all repetitions of either the general advice or advice offered in other sections.<\/p>\n [26. Even if your preferred AI doesn’t remember your preferences from one session to the next, you can force this behavior with a thread explicitly for ‘random questions’. This thread can be considered disposable – whenever necessary, just start a new one, telling the AI that it’s a continuation of the previous thread.]<\/p>\n 1. “Here is the situation, what are the logical steps for the antagonist to take next?” is a perfectly good application of AI because it plays to the AI’s strengths. “I need a villain with a plan” is a perfectly rotten use of AI.<\/p>\n 2. Ask it to guess where the story will go next, based on the most commonly used tropes and themes within the genre – then make sure it goes somewhere else. Ask it to analyze a character’s actions and determine his or her underlying motivation.<\/p>\n 3. Have it brainstorm ideas with you – then treat its ideas as inspiration for your own. And tell the AI your idea, ask it to find faults and flaws, and how the idea could be enhanced. Get it to do research for you, the more tedious the better – then verify and validate what it offers you; you will still save time.<\/p>\n 4. Always remember – YOU do the writing. IT helps with plotting. Treat it like a very clever 10 year old who is helping out – very well read, but only knows the most common answers.<\/p>\n From Gemini:<\/p>\n 5. Run the “Pre-Mortem” Simulation: Once you and the LLM have finalized a complex narrative arc or a mystery plotline, ask it to run a simulation against the player characters: “Assume the players completely ignore the main clue and do the most unexpected thing possible. Based on our established setting rules, what are three logical ways the world reacts to keep the momentum going?”<\/em>\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 6. Building on 1 & 5: Ask it for the worst plausible outcome, not just the next logical step. Point 1 above gets you a good “what happens next” – but ask separately, “what’s the most anticlimactic or contrived way this could play out?” It’s often better at spotting weak, cliche, or dues-ex-machina outcomes than at avoiding them unprompted, so use it as a pitfall detector for your own plotting, deliberately, rather than hoping it self-corrects.\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 7. Ask what the characters don’t know.<\/p>\n Rather than focusing only on what should happen next, ask the AI to identify misconceptions held by different characters. Stories are often driven more by incomplete or incorrect information than by objective events.\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral)<\/p>\n 8. Run “parallel universe” simulations.<\/strong><\/p>\n Ask: “How would this plot change if [key assumption] were false? Generate 3 alternate timelines.”<\/em> Doing so forces the AI to stress-test your narrative against hidden dependencies. [and assumptions you<\/em> have been making without realizing it!]<\/p>\n 9. Use “reverse outlining” for coherence checks.<\/strong><\/p>\n Paste a draft scene or plot arc and ask: “List the cause-and-effect chain in this narrative. Flag any gaps or illogical jumps.”<\/em><\/p>\n AIs that are strong in logical analysis, like Mistral or ChatGPT, perform better here.\n<\/ol>\n<\/blockquote>\n Qwen also suggests,<\/p>\n 10. Ask it to find your Chekhov’s Guns. After outlining a scenario or arc, prompt: “List every element I’ve introduced that a player\/reader will expect to become significant later. Flag any that I appear to have forgotten or left dangling.”<\/em> AIs are quite good at tracking introduced-but-unresolved details across a long brief, because that’s fundamentally a pattern-matching task rather than a creative one. This is different from Vibe’s point 9 (reverse outlining for cause-and-effect logic); this is specifically about promised-but-undelivered narrative elements. Nothing breaks reader\/player trust faster than a mysterious locked door in session one that nobody ever opens.\n<\/ol>\n<\/blockquote>\n Meta adds,<\/p>\n 11. Use the “off-screen clock”. Prompt: “While the players do X, what are 2 things the antagonists\/faction are doing that the players won’t see yet but will feel the consequences of in 2 sessions?” LLMs are great at foreground plot, terrible at simulating a living world unless explicitly asked to run the background simulation. This forces it.<\/p>\n 12. Request a “player-facing vs GM-facing” version. For any twist or reveal, ask: “Write the version the players should believe now, and the version I as GM should know is actually true, and tell me what evidence would let them recognize the shortfall in their understanding and how that could lead them to the GM’s version in whole or in part.” It makes the AI keep secrets properly layered instead of leaking them into player-facing text.\n<\/ol>\n<\/blockquote>\n Deepseek suggests also,<\/p>\n 13. Ask for the player-facing version of the wrong<\/em> answer. Instead of “what do the players know?” ask: “What would be a plausible but incorrect conclusion the players could draw from the evidence they have, and what clues would support that incorrect conclusion? [What clues contradict it if and when they are noticed?]”<\/em> This is more useful than asking for the correct answer disguised as mystery, because it gives you the red herrings pre-built. The players will generate wrong theories anyway – this lets you prepare for them and make them feel earned rather than arbitrary.<\/p>\n 14. For mystery plots, ask the AI to identify the minimum evidence required to solve it. Prompt: “Given this mystery, what is the smallest set of clues a player must find to reach the correct conclusion, and what is the logical path connecting those clues?”<\/em> This forces the model to distinguish essential from decorative information, and it gives you a scaffold for pacing discovery. Too often, mystery plots generated with AI are over-stuffed with clues because the model defaults to abundance rather than parsimony.\n<\/ol>\n<\/blockquote>\n Cohere advisess,<\/p>\n 15. Use “reverse outlining” to check coherence. Paste a draft scene and ask: “List the cause-and-effect chain. Flag any gaps or illogical jumps.” This surfaces structural problems before they become entrenched.\n<\/ol>\n<\/blockquote>\n 1. It’s best to have a general concept of the plot before you start, but not to be fixed in stone any more than necessary – let the plot evolve to match what the characters created are bringing to the table. DO ask the LLM to assess the impact of a character on the plot, and how best to showcase the unique qualities of the character.<\/p>\n 2. As a general rule, 2-4 alternatives that you can pick from are better than one offering. And don’t be afraid to blend and mix-and-match in a second iteration – “I like this element of choice A but not the rest; I like everything about choice B except this element – let’s swap out the bit I dislike from B for the bit I like from A”.<\/p>\n 3. Describe the story role that the character is to fill, then ask for general character concepts that could fill that role.<\/p>\n 4. If you want to avoid cliches and tropes, tell it that. If you want to lean into generic types, tell it that instead. In other words, tell it HOW the character is to fill the story role in question. For secondary \/ minor characters expected to only appear once, I tend to go in the latter direction; for featured and\/or recurring characters, I want more depth and original thinking.<\/p>\n 5. Specify any required end-status for the character. If there are no such requirements, add asking the LLM to outline the possibilities and how best to handle them to the <\/p>\n From Gemini:<\/p>\n 6. Prompt for Internal Contradictions: Generic characters are boring. When generating NPCs, explicitly ask the LLM to give them one major internal contradiction. A prompt like “Give me a loyal imperial captain who harbors a secret, deeply hypocritical vice”<\/em> yields far more compelling campaign interactions than simply asking for a “flawed captain.” [Good advice at all times, actually].\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 7. Ask for [or Offer] a sample of dialogue, not just a description. A backstory and personality summary tells you what a character is; three lines of how they’d actually talk in a tense negotiation or a casual moment tells you how to run them at the table. Voice is the thing that’s hardest to invent on the fly during a session – get it settled in advance.\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 8. Give every important character competence.<\/p>\n When generating characters, explicitly ask the AI what the character is genuinely good at – not merely what they believe they’re good at. Competence creates opportunities for interesting stories far more reliably than piling on flaws.\n<\/ol>\n<\/blockquote>\n [It probably works well in the other direction as well – “What is the character not good at but thinks that they excel in, and how do they explain failures to themselves?”]<\/p>\n From Vibe (Mistral):<\/p>\n 9. Assign “archetype weights”.<\/strong><\/p>\n To avoid cliches, use a sliding scale for traits: “Generate a rogue NPC with: 60% trickster, 30% scholar, 10% brute. Show how these percentages manifest in their behavior.”<\/em> This forces the creation of nuanced hybrids over stereotypes. [Death to the cardboard cutouts!]<\/p>\n 10. Request “conflict hooks”.<\/strong><\/p>\n Instead of just backstory, ask: “What are 3 secrets this character would die to protect, and who else knows them?”<\/em> Instant plot hooks tied to the character.\n<\/ol>\n<\/blockquote>\n [An alternative for #10: “What are three causes this character would support, what activities would they perform to support each cause, and how far would they go to support the cause?” – sometimes you don’t want plot hooks, you just want the character to be doing something instead of being frozen in place waiting for a PC prompt.]<\/p>\n Qwen says:<\/p>\n 11. Ask for the character’s “unremarkable Tuesday.” Instead of (or in addition to) backstory and conflict hooks, prompt: “Describe what this character does on a completely ordinary day when nothing plot-relevant happens. What do they eat, fidget with, avoid, look forward to?”<\/em><\/p>\n Backstory tells you who the character was; plot hooks tell you what the character does to others. The mundane routine tells you who they are when no one is watching, and it gives you exactly the small behavioral details – the way they stir their coffee too long, the route they take to avoid a particular street – that make a character feel like a person rather than a quest-dispensing kiosk. This complements Claude’s point 7 (dialogue samples) by covering behavior<\/em> rather than voice.<\/em>\n<\/ol>\n<\/blockquote>\n From Meta:<\/p>\n 12. Give the character a private theory about the world that is wrong. Prompt: “What does this NPC believe to be true about how the world works that is factually incorrect, and how does that belief shape their daily decisions?” Wrong theories produce more consistent, playable behavior than lists of quirks, because every choice can be filtered through the error.<\/p>\n 13. Ask for what annoys them, not [just] what motivates them. Motivation prompts produce grand, noble answers. Annoyance prompts produce usable table behavior: “What small thing makes this NPC immediately irritated, and how do they try to hide that irritation?” You will roleplay irritation far more often than you will roleplay lifelong ambition.\n<\/ol>\n<\/blockquote>\n Deepseek adds:<\/p>\n 14. Ask what the character would never say. Dialogue samples (point 7 from Claude) tell you the character’s voice. Asking “What is one thing this character would never admit, and how would they deflect if pressed on it?”<\/em> tells you [something about] their emotional architecture [it might not be the whole story]. The deflection pattern is the thing you’ll actually roleplay at the table – most PCs don’t drill down to deep secrets, but they do notice when an NPC gets evasive.<\/p>\n 15. Ask for the character’s self-justification. Prompt: “When this character does something morally questionable, what internal story do they tell themselves to feel justified?”<\/em> This is more useful than a simple “alignment” or “flaw” because it gives you their internal logic, which produces consistent behavior across a wide range of situations. A character with a self-justification framework is easier to improvise than a character with a list of traits.\n<\/ol>\n<\/blockquote>\n Cohere added a number of points – but they were all repetitions of either the general advice or advice offered in other sections.<\/p>\n 1. Supply it with the image of a room. Ask it to describe the room – then you rephrase that description to compress it and add atmosphere. Ask it to describe what’s most likely to be in the room but that isn’t depicted in the image. Ask it what the room says about the owner’s personality, and if that doesn’t fit with the personality of the character, ask it how the room’s appearance should be changed to better reflect the owner you have in mind. Always have it feed off your prompts and first thoughts – instead of letting it do your thinking for you. Treat it as a specialist collaborator, not a co-author – that’s how you will get the most out of it, and how your story will get the most out of it.<\/p>\n From Gemini:<\/p>\n 2. The Five-Senses Pass: LLMs default heavily to visual descriptions. Once it gives you a description of a location, hit it with a follow-up prompt: “Rewrite this description, but completely remove all visual adjectives. Focus exclusively on the sounds, textures, ambient temperatures, and smells of the space.” This forces the AI to provide the exact kind of atmospheric sensory details that make theater-of-the-mind gaming memorable. [Make sure you manually edit \/ compress this output, however. See the Secrets Of Stylish Narrative<\/a> series.]\n<\/ol>\n<\/blockquote>\n From Claude:<\/p>\n 3. Ask it to re-describe the space from an unusual point of view – a character who’s blind, a creature far smaller or larger than human scale, someone who’s never seen the technology\/magic in the room before. This surfaces details and framing that a straightforward description misses, and it’s a different exercise from the Five-Senses Pass above (that strips a sense; this changes the observer).\n<\/ol>\n<\/blockquote>\n From ChatGPT:<\/p>\n 4. Ask how people actually use the space.<\/p>\n A believable location isn’t just architecture. Ask what an ordinary visitor notices after five minutes, what the owner has gradually stopped noticing, and what evidence there is of everyday use. These small details often make a location feel lived-in.\n<\/ol>\n<\/blockquote>\n From Vibe (Mistral)<\/p>\n 5. Ask for “functional affordances”.<\/p>\n “What can a character realistically DO in this space? List 5 interactive elements (eg, ‘the chandelier can be swung from’, ‘the desk drawer is locked’).”<\/em>. Helps turn static descriptions into dynamic gameplay tools.\n<\/ol>\n<\/blockquote>\n From Qwen:<\/p>\n 6. Ask how the space shows its history. Prompt: “What wear, repairs, modifications, stains, or accumulated clutter in this space tell the story of the people who have used it over time?”<\/em><\/p>\n A room described only in its present state feels like a stage set. A room where the paint is newer around the door-frame because someone moved a bookcase, where the third stair creaks and someone has wedged a folded card under it, where there’s a faded rectangle on the wall where a portrait used to hang – that room has duration.<\/p>\n This is related to ChatGPT’s point 4 (how people use the space) but focuses on temporal layering rather than present-tense activity [an important distinction – the two are complimentary].\n<\/ol>\n<\/blockquote>\n Meta offers:<\/p>\n 7. Do a “looting pass”. After you have a good description, ask: “If a curious, slightly kleptomaniac PC searches this room for 10 minutes, what 5 small, non-valuable but story-rich objects would they find, and where exactly are they tucked away?” This turns atmosphere into interactive inventory without you having to invent it, and is distinct from functional affordances – these are not tools to use, they are evidence that someone lived here.<\/p>\n 8. Ask for the maintenance question. “Who cleans this place, who repairs it, and what shortcut have they taken that they hope no one notices?” Every location has upkeep. The shortcut is where adventure lives.<\/p>\n [Alternatively, if no-one is providing maintenance, ask for what has broken and been repaired, what has broken and needs to be repaired, and what are the most visible signs of decay that are not already noted.]\n<\/ol>\n<\/blockquote>\n From Deepseek:<\/p>\n![\"\"](\"https:\/\/www.campaignmastery.com\/blog\/wp-content\/uploads\/2026\/08\/MLRv1-Bg3-SM.jpg\")
\n
The ethical use of AI<\/h5>\n
Three Flavors of AI<\/h3>\n
The Perplexity Puzzle<\/h5>\n
\n
\n
Comparing AIs<\/h5>\n
The Bottom Line:<\/h6>\n
General Advice<\/h3>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
Generative AI (Images) Advice<\/h3>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n2. Re-prompt only<\/em> for that region (eg, “a hand with 5 fingers, relaxed pose”).<\/p>\n
\n ▪ “Absence of X” (eg, “absence of modern technology”).
\n ▪ “Minimalist, only [Y]” (forces exclusion by focus).\n<\/ol>\n<\/blockquote>\n\n
\n
\n
\n
\n
\n
\n
\n
When it all goes sideways: Generative AI failure and recovery<\/h5>\n
\n
\n
\n
\n
When to stop using generative AI and composite<\/h5>\n
\n
\n
\n
\n
LLMs (Text) Advice<\/h3>\n
General LLM Advice<\/h5>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n ▪ Turn these into stat blocks:
\n ▪ Example 1: “Goblin, HP 10, AC 14, +2 dagger”
\n ▪ -> **Goblin | HP 10 | AC 14 | Atk: +2 dagger**
\n ▪ Example 2: “Ogre, HP 30, AC 12, +5 club”
\n ▪ -> **Ogre | HP 30 | AC 12 | Atk: +5 club**
\n ▪ Now do this one: “Troll, HP 45, AC 15, +7 axe”<\/p>\n\n
\n
\n
\n
\n
\n
\n
\n
Prepare a Briefing Document<\/h5>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n ▪ ### for subsections (eg, “Factions > The Iron Pact”).
\n ▪ #### for granular details (eg, “Factions > The Iron Pact > Leadership”).<\/p>\n
\n ▪ – Always prioritize [X] over [Y].
\n ▪ – If unsure, ask: “Did you mean [A] or [B]?”
\n ▪ – Never assume [Z].<\/p>\n\n
\n
\n
\n
\n
\n
Continue Chat Threads<\/h5>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n ▪ Feed select parts back to the AI (eg, “Here’s our discussion on magic items – now design a new one following these rules”).\n<\/ol>\n<\/blockquote>\n\n
\n
\n
\n
\n
\n
Usage as a plotting assistant \/ co-brainstormer<\/h5>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
Usage as a character generator<\/h5>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
Usage to refine location ideas \/ descriptions<\/h5>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n