Feed any AI output back to a model and reconstruct the prompt behind it — then improve on it. Works on images, video and text.
READ8 min
WORDS1,513
TYPEGUIDE
CHECKED25 AUG 26
TL;DR — THE SHORT VERSION
Reverse prompting asks a language model to rebuild the prompt behind an AI-generated output. It is a fast way to learn prompt structure, but the model is inferring, not measuring.
Paste a reverse prompt, then the output. The model breaks it into visual elements, readable settings, a reconstructed prompt and an improved version.
The look is recoverable. Composition, lighting, colour grade, apparent lens, styling and mood can be inferred well.
Generation settings are not. CFG scale, sampler, step count, seed and LoRA leave no reliable trace, so a specific answer about them is invented.
Check the file first. Many generated images carry their real prompt in metadata; guess only when it has been stripped or re-encoded.
Expect weak results on composites. Retouched, hand-edited or multi-tool work had no single prompt behind it.
Most people guess at what makes a good prompt. Reverse prompting shows you. Every AI-generated thing has a prompt architecture underneath; deconstruct it and you learn faster than any tutorial teaches.
Three reverse prompts, for images, video and text. Feed one an image description, a video scene or an output: it infers the prompt architecture behind it, marks what it cannot recover, then writes an improved prompt.
WHY REVERSE PROMPTING?
Most people try to guess what makes a great prompt. Reverse prompting shows you exactly. Every AI-generated image, video, or text has a hidden prompt architecture — deconstruct it, and you learn faster than any tutorial.
✦ Works on images
✦ Works on videos
✦ Works on LLM outputs
✦ Any AI model
PRO TIP: Use it on output from any model you like — FLUX, Midjourney, Seedance, Higgsfield, ChatGPT Images 2.5, Runway or an LLM. The reconstruction is an inference either way.
TL;DR — HOW REVERSE PROMPTING WORKS
01Copy the Universal Reverse Prompt below using the button
02Paste into GPT-6, Claude, or Gemini as your first message
03Describe or paste the AI output you want to reverse-engineer
04Get a full 5-phase deconstruction: visual breakdown → technical params → reconstructed prompt → negative prompt → optimized version
05Use Phase 5 (Optimized Version) as your new, improved prompt
PH.2Technical Parameters — what is readable (ratio, resolution) and what is not (CFG, sampler, seed, LoRA)
PH.3Reconstructed Prompt — a best inference of a prompt that would produce this output
PH.4Negative Prompt — a suggested one, labelled as a suggestion; what was actually excluded cannot be read from the output
PH.5Optimized Version — an improved prompt that exceeds the original
WHAT REVERSE PROMPTING CANNOT TELL YOU
Read this before trusting any output above. A language model reading an image — or your description of one — is inferring, not measuring. Some things it infers well. Some it cannot know at all, and will invent rather than admit.
Genuinely recoverable: composition, lighting direction and quality, colour grade, apparent focal length and depth of field, styling, mood, era cues, and the shape of the prompt that would produce them. This is the part worth doing, and it is most of the value.
Not recoverable from an image: CFG scale, sampler, step count, seed, scheduler, or which LoRA was applied. None of these leave a reliable signature. Ask a model for them and you will get a confident, specific, invented answer — the same failure mode described in why AI makes things up. The prompts on this page are written to refuse rather than guess; if you use your own version, build in the same refusal.
Check the file before you guess at it. Many generated images carry the real prompt in their metadata — a PNG text chunk or EXIF field — put there by the tool that made them. Reading that is exact and free. Guess only when the file has been stripped, screenshotted, or re-encoded, which social platforms do routinely.
Where it degrades: composites, heavily retouched work, hand-edited output, and anything that passed through several tools. There was never one prompt behind those, so a reconstruction will be fluent and wrong.
WHAT AN IMAGE GIVES BACK
Check the file’s metadata first. Without it, infer only what leaves a visible trace.
RECOVERABLE · INFER IT
Yes: Composition
Yes: Lighting direction and quality
Yes: Colour grade
Yes: Apparent focal length and depth of field
Yes: Styling, mood and era cues
Note: Less so for composites and multi-tool work
NOT RECOVERABLE · REFUSE
No: CFG scale
No: Sampler and step count
No: Seed and scheduler
No: Which LoRA was applied
Reasoning — summarises this page’s box on what reverse prompting cannot tell you. Page checked 25 Aug 2026.
TAKEAWAY
Treat a reconstructed prompt as a hypothesis. Use it to learn how a result was structured, and distrust any specific number a model gives for settings an image cannot reveal.
★ MASTER REVERSE PROMPT — UNIVERSAL
Universal Reverse Prompt — Extracts Everything From Any AI Output
A general-purpose reverse prompt. Feed it any AI output — it breaks down the prompt architecture, names the techniques it can see, marks the settings it cannot recover, and writes an improved prompt.
You are an expert AI prompt analyst and reverse engineer. Your task is to fully deconstruct any AI-generated content I provide and extract the complete prompt architecture.
When I give you an AI output (image description, video scene description, or text), analyze it and provide:
PHASE 1 — VISUAL/CONTENT BREAKDOWN:
Identify: composition, subject positioning, lighting quality and direction, color palette, atmosphere, texture detail, perspective, focal length feel, and every significant visual element.
PHASE 2 — TECHNICAL PARAMETERS (label your confidence honestly):
Aspect ratio and approximate resolution are readable from the output itself — state those plainly.
Model family, CFG scale, sampler and any LoRA are NOT recoverable from an image or a description. Do not guess a number. For each one, answer in this form:
- AVOID INVENTING: if the value cannot be determined from what I gave you, write "not inferable" and stop. Do not supply a plausible-sounding number.
- IF THERE IS EVIDENCE: name the visual evidence first, then give a range rather than one value. Example: "heavy plastic skin and waxy highlights are consistent with high guidance — possibly CFG 9-12, low confidence."
- NEVER state a sampler name, step count, seed, or LoRA title as fact. These leave no reliable trace in an image.
PHASE 3 — RECONSTRUCTED PROMPT:
Write the prompt that most plausibly produced this output, as a best inference, not a recovered original. Include: subject description, lighting, camera specs, mood, style references, quality tags, and only the technical parameters Phase 2 found readable.
PHASE 4 — NEGATIVE PROMPT:
A negative prompt cannot be read from an output. Suggest one that would help reproduce this look, and label it "suggested, not recovered".
PHASE 5 — OPTIMIZED VERSION:
Improve upon the original. Write an enhanced version of the prompt that would produce superior results — higher quality, more cinematic, more detailed, more intentional.
Begin your analysis now. I will provide the AI output in my next message.
TEXT-TO-IMAGE REVERSE PROMPT
Image Reverse Prompt — Maximum Visual Detail Extraction
Engineered for reverse-engineering AI-generated images. Extracts every visual parameter — from pixel-level texture to macro composition choices — then rebuilds and supercharges the prompt for FLUX, Midjourney, SDXL, and Nano Banana Pro.
You are an expert visual analyst and AI image prompt engineer. Analyze the image I describe and reverse-engineer its complete prompt.
For the image I provide, extract:
1. SUBJECT: exact description of who/what, clothing, expression, pose, positioning
2. LIGHTING: light source type, direction, quality, color temperature, shadows
3. CAMERA: estimated focal length, aperture (depth of field), camera height and angle
4. COMPOSITION: framing, rule of thirds, negative space, foreground/background relationship
5. COLOR GRADE: overall palette, warm/cool balance, saturation, contrast level
6. STYLE: publication or aesthetic reference (Vogue, NatGeo, editorial, documentary)
7. TECHNICAL: quality level, sharpness, grain, any post-processing artifacts
Then produce:
— RECONSTRUCTED PROMPT: copy-paste ready, full detail, labelled as a best inference. Do not state CFG, sampler, step count, seed or LoRA; write "not inferable" for each.
— OPTIMIZED ENHANCED PROMPT: improved version that would exceed the original quality
TEXT-TO-VIDEO REVERSE PROMPT
Video Reverse Prompt — Scene & Motion Deconstruction
Built for text-to-video AI. Deconstructs camera movement, temporal consistency, scene progression, subject motion, and cinematographic language — then reconstructs and improves the complete video prompt.
You are an expert cinematographer and AI video prompt engineer. Analyze the video scene I describe and reverse-engineer its complete prompt in this bracket format.
For the scene I describe, extract:
[MOTION]: what is moving, how it moves, at what pace
[SUBJECT]: who/what, their consistent appearance, identity markers
[ENV]: full environment description, time of day, weather, atmosphere
[CAMERA]: shot type, camera movement, lens feel, frame rate
[LIGHTING]: all light sources, quality, color, dramatic intent
[STYLE]: director reference, cinematic language, film aesthetic
[DURATION]: estimated clip length
Then produce:
— RECONSTRUCTED BRACKET PROMPT: complete, in the bracket format above
— OPTIMIZED VERSION: enhanced prompt for superior temporal consistency and cinematic quality