Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/nidhinjs/prompt-master/llms.txt

Use this file to discover all available pages before exploring further.

Image generation AI and language models respond to prompts in fundamentally different ways. A language model rewards specificity about structure and reasoning. An image model rewards specificity about visual information — subject, style, lighting, composition, and parameters — presented in a particular order. Inject the wrong structure and you get a mediocre output no matter how well you described the image you wanted. Prompt Master’s four image and media templates are calibrated to the syntax, parameter conventions, and known failure modes of each generation platform. Detection is automatic: the template selection happens before you see the output, based on the target tool, the presence of an existing prompt or reference image, and whether the task is generating something new or editing something that exists.

Visual Descriptor

Visual Descriptor is the primary template for all text-to-image generation. It structures the prompt in the order that diffusion models weight most heavily: subject first, then style and mood, then technical parameters, then the negative prompt. Reordering these layers produces measurably worse results across all major generation platforms. Detection logic: Target tool is Midjourney, DALL-E 3, Stable Diffusion, Sora, Adobe Firefly, or any generative image/video AI, and the task is creating a new image or video from a text description (no reference image provided).
LayerWhat it carries
SubjectThe primary element — who or what the image is of, with specific descriptors
Style / MediumThe visual style, artistic medium, or photographic genre
Mood / AtmosphereThe emotional register — cinematic, melancholic, energetic, serene
LightingLight source, direction, quality — golden hour, neon, hard shadow, diffused
CompositionFraming, perspective, depth of field — wide shot, close-up, Dutch angle
Technical detailRendering quality descriptors — photorealistic, ultra detailed, 8K, RAW
ParametersTool-specific flags at the end — --ar, --v, --style, --cfg, steps, etc.
Negative promptWhat to exclude — only for tools that accept a separate negative prompt field
Example — Midjourney prompt generated by Prompt Master:
lone samurai standing in heavy rain at night, traditional armor, neon reflections on wet cobblestone street, cinematic lighting, dramatic shadows, fog, ultra detailed, photorealistic, shallow depth of field --ar 16:9 --v 6 --style raw
Negative prompt: cartoon, anime, watermark, blurry, oversaturated, flat lighting, daytime, modern clothing, text
Subject-first ordering is non-negotiable. Diffusion models weight tokens by position — tokens at the start of the prompt have stronger influence on the output. Putting style or lighting before the subject shifts the model’s attention away from what you actually want in the image.
Example — DALL-E 3 prompt generated by Prompt Master:
A macro photograph of a single raindrop balanced on the tip of a pine needle, early morning forest background softly blurred, cool blue-green tones, rim lighting from behind, high contrast, photorealistic, extreme detail on the water droplet surface, DSLR quality

Reference Image Editing

Reference Image Editing is used when the task is modifying an existing image rather than generating a new one from scratch. Prompt Master detects the intent automatically — if your request includes an existing image or explicitly describes editing something that already exists, this template is selected over Visual Descriptor. Detection logic: The user provides a reference image URL or file, or the request uses editing language (“change the background,” “remove the object,” “replace the lighting,” “make the subject wearing a different outfit”). Prompt Master identifies edit intent vs. generate intent and routes accordingly — you do not need to specify which template to use.
ComponentWhat it carries
Reference descriptionA factual description of what’s in the reference image — grounding the model’s understanding
Edit instructionThe specific change to make, described with precision
Preserve listElements of the original image that must not change
Style matchWhether to match the original’s style or introduce a new one
ParametersTool-specific editing flags — --iw (image weight) for Midjourney, inpainting mask notes for SD
Example — Reference image editing prompt generated by Prompt Master:
Reference image: a product photo of a white ceramic coffee mug on a wooden table, natural window light from the left, clean white background.

Edit: Replace the white background with a dark, moody coffee shop interior — exposed brick wall, warm Edison bulb lighting from the right, shallow depth of field. The mug and table surface must remain unchanged.

Preserve: The mug shape, color, and texture. The wooden table surface. The natural left-side lighting on the mug itself.

Style match: Shift from clean product photography to editorial lifestyle photography. Match the warm amber tones of the new background environment.

--iw 2 --v 6
Edit vs. generate detection is automatic. If you paste a prompt like “change the background of this image to a forest,” Prompt Master routes to Reference Image Editing. If you say “generate an image of a person in a forest,” it routes to Visual Descriptor. You never need to specify the template.

ComfyUI

ComfyUI prompts have a different structure from standard diffusion prompts because ComfyUI is node-based — the positive and negative prompt are separate inputs to specific nodes, and different parts of the workflow (base checkpoint, refiner, ControlNet, LoRA) can accept different prompts. A flat prompt string pasted into ComfyUI will work, but a prompt written specifically for the node architecture produces significantly better results. Detection logic: Target tool is explicitly ComfyUI, or the user mentions a node-based workflow, SDXL with a refiner, ControlNet conditioning, or a LoRA stack.
ComponentWhat it carries
Positive prompt (base checkpoint)Main subject and composition — written for the base model’s training data
Positive prompt (refiner)Detail and quality descriptors — written for the refiner’s upscaling pass
Negative promptExplicit exclusions — comma-separated, targeted at the base checkpoint’s known failure modes
ControlNet noteIf ControlNet conditioning is in use: what type of conditioning and what it should enforce
LoRA weightsIf LoRAs are active: name, trigger word, and recommended weight
Sampler settingsRecommended steps, CFG scale, and sampler for the described output
Example — ComfyUI prompt set generated by Prompt Master:
[Positive prompt — base checkpoint (SDXL 1.0)]
portrait of a botanist in a Victorian greenhouse, surrounded by exotic plants and glass panels, natural diffused daylight, warm green tones, linen clothing, focused expression, medium shot

[Positive prompt — refiner (SDXL Refiner 1.0)]
highly detailed skin texture, sharp fabric weave on linen jacket, crisp glass reflections, photorealistic, 8K, film grain

[Negative prompt]
cartoon, illustration, painting, watermark, text, signature, deformed hands, extra fingers, blurry, low resolution, oversaturated, purple tint, harsh shadows, plastic skin

[ControlNet]
OpenPose — enforce upright standing pose, arms relaxed at sides

[Sampler settings]
Steps: 30 | CFG: 7.5 | Sampler: DPM++ 2M Karras
Base: 0–0.8 | Refiner: 0.8–1.0

Prompt Decompiler

The Prompt Decompiler is a meta-template — it doesn’t generate a new prompt, it operates on an existing one. When you paste a prompt into Prompt Master and ask to break it down, adapt it, simplify it, or split it into components, the Prompt Decompiler activates. It identifies the framework the original prompt was using, labels its components, diagnoses problems, and produces the requested transformation. Detection logic: The user pastes an existing prompt and uses any of the following intent signals: “break this down,” “explain this prompt,” “simplify this,” “adapt this for [different tool],” “split this into two prompts,” “what framework is this using,” or pastes a prompt with no other context (Prompt Master defaults to decompilation mode for bare prompt pastes).
Output typeWhat it produces
DecompositionLabels each part of the original prompt with its function and the framework it belongs to
DiagnosisIdentifies anti-patterns, token waste, missing components, and conflicts
SimplificationStrips non-load-bearing language while preserving the prompt’s effective instructions
AdaptationRewrites the prompt for a different target tool using the correct template for that tool
SplitDivides a bloated single prompt into two or more focused prompts
Example — Prompt Decompiler input and output: Input (pasted by user):
Please act as an expert marketing professional with years of experience in the field and write me a really good comprehensive detailed blog post about content marketing strategies for small businesses that covers all the important topics and is very helpful and informative and engaging for the reader and also make sure it's SEO-friendly.
Prompt Decompiler output:
Framework detected: RTF (malformed — missing Format component, Role is weak)

Decomposition:
- Role: "expert marketing professional with years of experience" → weak (no specialization, no behavioral signal)
- Task: "write a blog post about content marketing strategies for small businesses" → valid but underspecified (no word count, no angle, no target reader)
- Format: missing → model will choose arbitrarily

Anti-patterns found:
- Stacked quality adjectives: "really good," "comprehensive," "detailed," "very helpful," "informative," "engaging" — none of these constrain the output. Remove all.
- "Covers all the important topics" — unbounded scope signal. Replace with a specific topic list.
- "SEO-friendly" — not actionable without specifying keywords, meta description, or H-tag structure.

Simplified and corrected prompt (RTF):

You are a content strategist who has built organic traffic channels for 50+ small businesses using primarily SEO-driven blog content.

Task: Write a 1,200-word blog post titled "5 Content Marketing Strategies That Work for Small Businesses With No Ad Budget." Target reader: small business owners with 1–10 employees who handle their own marketing. Include one H2 per strategy, a 155-character meta description, and a CTA in the final paragraph pointing to a free content calendar template.

Format: Markdown. H1 title, 5 × H2 sections, closing CTA paragraph. No introduction fluff — open with the first strategy. Reading level: Grade 8.
The Prompt Decompiler is also useful for adapting prompts across tools. If you have a well-performing Claude prompt and want to use the same task with Midjourney or GPT-4o, paste it and ask: “Adapt this for [target tool].” Prompt Master will identify what needs to change for the new tool’s syntax and re-template it accordingly.

Image Template Detection Summary

SignalTemplate selected
Target: Midjourney, DALL-E, SD, Sora, Firefly + no reference imageVisual Descriptor
Reference image provided, or editing language detectedReference Image Editing
Target: ComfyUI, or node-based workflow / LoRA / ControlNet mentionedComfyUI
Existing prompt pasted + break down / adapt / simplify / split intentPrompt Decompiler
Existing prompt pasted with no other contextPrompt Decompiler (default)
Detection is automatic. You never need to say “use Visual Descriptor” or “use the Prompt Decompiler.” Prompt Master infers the correct template from what you’ve provided and what you’ve asked for.

Build docs developers (and LLMs) love