69 templates, each a real graph you can open with one click — filter by category or search by what it does, then run it on your own input.
69 of 69 templates
Document
Extract text from a PDF document and summarise it with an LLM in 5 bullet points. Demonstrates document.extract_text chained into a model step.
2 steps · Extract text from document
Research
Run a web search and summarise the top results with an LLM, including citations. Demonstrates web.search chained into a model step.
2 steps · Web search
Video
Extract a frame from a video at 1 second and resize it to a 320×180 thumbnail. Demonstrates video.extract_frame chained into image.resize.
2 steps · Extract single frame, Resize image
Content
Turn a long-form article into a numbered social media thread ready to post. Single model step demonstrates a focused rewriting/repurposing prompt.
1 step
Fetch the given URL via the http step, then have an LLM extract the key points as a focused, bulleted research brief. Demonstrates http -> model chaining for extraction/summarisation.
2 steps
Fetch the given URL via the http step, then have an LLM rewrite it into an engaging blog post in the style you specify. Demonstrates http -> model chaining via ${steps.X.output} bindings.
Localization
Translate a single piece of text into several languages in parallel via foreach. Demonstrates ${item} inside a foreach.step subtree with items bound from ${input.languages}.
Cut the subject out of a product photo with a purpose-built background remover, then place and retouch it on a clean studio backdrop. Two image models from two vendors, chained by feeding the first step artifact into the second step input.image.
Generate a character front view, then derive the side and back views from it so the design stays consistent. Each view is edited from the previous one rather than generated afresh.
Recompose one source image into wide, square and vertical thumbnail variants at the same time. The three media steps read only the source image, so the engine sees they are independent and runs them together — the whole set costs one round of wall-clock rather than three.
Data
Compute real totals and averages from a CSV inside a sandboxed Python step, then have an LLM write the summary from those computed figures rather than from the raw rows. Demonstrates code.run_python feeding a model step.
2 steps · Run Python code
Write a report with an LLM, render it to PDF with pandoc, generate a cover image, then zip both into one downloadable bundle. archive.pack takes each file as an entries[].sourceUrl binding, so the archive is assembled from two artifacts made earlier in the same run.
4 steps · Convert document format, Create archive
Audio
Take the first two minutes of an episode, re-encode it to a compact mp3, transcribe it, and write show notes from the transcript. The transcription step is a media step naming a real STT model, so it publishes text rather than an artifact — bind it as output.text, not outputRef.
3 steps · Multi-op audio pipeline
Cut the product out of a photo with a purpose-built background remover, drop it onto a clean white sheet, put that sheet on a fixed 1400px canvas, add even margins and stamp your logo in the corner. Five artifacts chained end to end: a model output into four sharp operations, plus a separate branch that shrinks the logo. The resize is what makes the margins and the mark mean the same thing on every photo — without it a 140px border is invisible on a 6000px shoot. Supply a PNG logo with a transparent background — it is composited with over, so its own background would show.
5 steps · Flatten alpha, Resize image, Add padding, Composite (overlay) images
Generate a brand mark with an image model, shrink it, then stamp it onto wide, square and vertical crops of your photo — all three crops at once. The mark is a real artifact: each crop is an image.pipeline that resizes the photo and then composites the mark, binding it through params.operations, and multiply blending keeps the photo visible around it. The three crops read the same two things and nothing of each other, so they run together.
3 steps · Resize image, Multi-op image pipeline
Turn rough notes into a Marp slide deck, render it to PDF, read the rendered PDF back out as text, and write a launch post from that text. The handout is extracted from the deck artifact itself, so it can never drift from the slides that were actually rendered.
3 steps · Render markdown slides, Extract text from document
Generate a short video and an original music bed from two prompts, then mux them together with ffmpeg. Two model artifacts land in ONE capability input: video.add_audio takes video_url and audio_url, and both are bound as outputRef. Start this one with mode async — its video step declares a 300 second timeout, and sync runs cap a pipeline at 120.
2 steps · Add audio to video
Turn exactly nine images into one 3x3 contact sheet. A foreach squares every image to 400x400, then a single image.pipeline stretches the NINTH tile to 1200x1200 as the canvas and composites all nine onto it, one per gravity position — the canvas ends up fully covered, so which tile it was made from does not show. Nine is not a placeholder: image.composite has nine gravities and no x/y offsets, which is exactly a 3x3 and nothing wider. Fewer than nine images is refused when the sheet step validates its input; more than nine is refused before the loop runs at all.
Scale a video down once, pull nine frames from its first nine seconds, and composite them into one 3x3 grid. Nine is the ceiling, not a choice: image.composite has nine gravity positions and no x/y offsets. The scale step runs first and alone because a video.pipeline that also extracts a frame silently discards it. Costs eleven transform operations — one scale plus nine frame extracts that each re-fetch the scaled clip, then the grid — so expect a credit line of eleven, not one. Needs a clip of at least nine seconds; a shorter one fails at the first frame past its end.
3 steps · Multi-op video pipeline, Extract single frame, Multi-op image pipeline
Turn a photo into clean black line art and lay it out for A4 print. The image model does the conversion; the two capability steps after it do the things a model is unreliable at — `image.flatten` paints the transparency white so a printer does not render it black, and `image.resize` with `fit: contain` places the drawing inside a 2480x3508 page (A4 at 300 dpi) without cropping it. Both run on the artifact, not on a re-generated picture, so the line work is untouched.
3 steps · Flatten alpha, Resize image
Cut the first fifteen seconds out of a landscape video, crop it to a vertical 9:16 window, and stamp a generated brand mark into the corner. THE CROP IS COMPUTED FOR A 1920x1080 SOURCE: a 9:16 window 1080 tall is 607 wide, centred at x=656. A source of any other resolution needs different numbers — video.crop takes absolute pixels and has no aspect-ratio form, so it will crop the wrong region rather than fail. A source shorter than fifteen seconds fails at the trim.
5 steps · Resize image, Trim video, Crop video frame, Overlay watermark
Read a reference photo into a written character sheet, then render that character somewhere new. The model looks at the photo itself and pins hair, build, wardrobe and palette in words first, and the render is given both the reference image and those words. Directing the image model from the photo alone tends to keep the pose and lose the person.
Search the web for the event, have a model pull out the facts that belong on a poster, then art-direct the poster around them. The search step exists so the poster is about a real event rather than a plausible one — the model is told to use only what the results contain and to leave a detail out rather than invent it. The date and venue are still worth checking: the image model sets the text, and it is the least reliable part of any generated poster.
3 steps · Web search
Straighten a photographed whiteboard into a clean diagram, then make it presentation-ready. The image model redraws the boxes, arrows and handwriting as crisp black-on-white; `image.flatten` paints any transparency white so it does not print black; `image.resize` with `fit: contain` fits it into a 1920x1080 slide without cropping an arrow off the edge. Handwriting is transcribed by the model, so check the labels — that is the step most likely to be wrong.
Put top and bottom meme captions onto a photo in the classic heavy white outlined face. The text is rendered by the image model rather than composited, because there is no text operation in the catalogue — which also means the wording can come out slightly off, and re-running is the fix. Captions are passed verbatim in quotes so the model sets them rather than rewrites them.
Turn a pencil sketch or line drawing into a finished render, keeping the drawn geometry and changing only material, light and finish. The prompt is written to constrain rather than to inspire — an image model given a sketch and an open brief tends to redraw the subject, which loses the design the sketch was made to fix.
Cut three five-second windows out of a long recording and join them into one fifteen-second reel. THE WINDOWS ARE FIXED at 0-5, 20-25 and 40-45 seconds, so the source must be at least 45 seconds long — a shorter one fails at whichever trim runs past its end rather than returning a shorter reel. `video.concat` takes a list of at least two videos in `input.videos`, which is why the three cuts are separate steps rather than one pipeline.
2 steps · Trim video, Concatenate videos
Cut the product out of its photo, generate a billboard scene, and composite the cutout into the centre. The cutout is a real artifact with an alpha channel, which is why it can sit on the backdrop without a white box around it — this is the difference between compositing and asking an image model to "put it on a billboard", which redraws the product. The cutout is scaled to 900px before compositing so its size means the same thing on any source photo.
Restyle a portrait into the 1950s, 1970s, 1990s and 2010s, then lay all four into one 2x2 contact grid. Each decade is edited from the ORIGINAL photo, not from the previous decade, so four edits do not compound into a fifth face. The grid is 1024x1024 with 512-wide tiles placed on the four corner gravities: `image.composite` has nine gravity positions and no x/y offsets, so 2x2 is exactly what it can express and the two sizes have to agree.
Break a recipe into discrete steps and draw each one in a consistent style. The model returns a JSON array of step descriptions and the loop draws one image per entry, three at a time. `max_iterations` is 6 and the prompt asks for at most six steps — the cap is what stops a long recipe from turning into an unbounded bill, and it truncates rather than fails, so a longer recipe silently loses its tail.
3 steps
Write an icon brief, render it once, then export it at 512, 256 and 64 pixels. The three exports are capability steps on the same artifact rather than three generations, so the icon is identical at every size — asking an image model for "the same icon, smaller" gets you a different icon. `fit: contain` keeps the whole mark inside the square instead of cropping into it.
3 steps · Resize image
Write three competing title treatments, then art-direct a 16:9 thumbnail around the strongest one and export it at exactly 1280x720. Distinct from `video-thumbnail`, which pulls a frame out of an existing video — this one designs a thumbnail for a video that may not be shot yet. The resize is `fit: cover` because a thumbnail slot is a fixed rectangle and letterboxing it wastes the frame.
Put a product photographed anywhere onto a clean seamless studio backdrop, keeping the product itself untouched. The prompt constrains hard because the failure mode is not a bad backdrop — it is a redrawn product that no longer matches what you sell.
A framed trading card built around a subject: art, border, title plate and a rarity treatment. The title is set by the image model, so check the spelling — text is the least reliable part of any generated card, and the catalogue has no text operation to fix it with.
Three icons of one character or theme, cut out and laid across a single strip so they can be checked as a set. THE STRIP IS 1200x400 WITH 400px TILES on the west, center and east gravities: `image.composite` has nine gravity positions and no x/y offsets, so a three-across strip is exactly what it can express and the two sizes have to agree.
Break a scene into panels, then draw the whole page in one pass. One generation, not one per panel: a comic page is read as a single composition, and six separately drawn panels do not share a horizon, a palette or a character's face.
Write the cover lines, art-direct the cover around them, then export at a fixed print ratio. One tool for both jobs the dump lists separately — a magazine cover and a book cover differ in the words, not the pipeline. The model sets the type, so proof the title before sending it anywhere.
Blend two animals into one plausible creature and render it as a naturalist study. Generated from text rather than edited from a photo, because a fusion has no source image to preserve — asking an editor to merge two animals tends to return the first one wearing the second.
Six expressions of one character on a single sheet, drawn from a reference photo or artwork. One generation rather than six: an expression sheet is judged on consistency, and six separate edits drift in face shape long before they drift in expression. Consolidates the dump's five near-identical emote and emotion-sheet flows.
Turn a flat floor plan into a furnished three-quarter view of the same layout. Walls, doors and room proportions are held; only the furnishing and the viewpoint are invented — a render that quietly moves a wall is worse than no render.
A single isometric tile rendered to drop straight into a game grid: fixed camera angle, transparent-ready edges, nothing bleeding past the diamond. The prompt pins the projection because "isometric" alone gets you a perspective drawing that will not tile.
Four distinct luxury advertising treatments of the same product, generated from the same photo so the product is constant and only the art direction moves. Four separate edits rather than one grid, because each is meant to be used on its own.
Explain a process, then draw it as a textbook schematic and keep the explanation alongside as the caption. Distinct from `technical-diagram`, which draws an object's parts — this one draws a sequence, with stages and arrows.
Re-dress an existing asset for a season or event without changing what it is. Silhouette, pose and readability are held so the asset still reads at icon size — the point is a variant, not a replacement.
A sheet of die-cut style stickers on one theme, each with the thick white keyline that makes a sticker read against any background. Generated as one sheet so the set is stylistically consistent; cutting them apart is a separate job the catalogue cannot do yet.
Turn a description of how something works into a clean labelled diagram, flattened and sized for a slide. The model writes the parts list first so the drawing has something to be faithful to; the labels are still rendered as image text, so they are the thing to check.
4 steps · Flatten alpha, Resize image
Pull the garments out of a photographed outfit and lay them out as a catalogue flat-lay, one item at a time, on white. Useful for building a product grid from lookbook shots. The person is removed; what they were wearing is not reinterpreted.
Lift hard shadows and blown highlights out of a photo while leaving the subject's own form intact. Distinct from a background removal: nothing is cut out, the lighting is rebuilt. Shadows that carry shape — the ones under a chin or inside a fold — are meant to stay, and the prompt says so.
Relight and restage a phone photo of a dish into something that belongs on a menu, without changing what is on the plate. Portion, garnish and plating stay as photographed — the model is told to change light, background and camera only, because a menu photo of a different dish is a refund.
A restrained skin and detail pass on a portrait: even the tone, keep the texture. The prompt asks for pores and fine lines to survive, because the default failure of every automatic retouch is a plastic face, and that is much harder to undo than a blemish.
Read a moodboard into a written style guide, then make something new that obeys it. The written middle step is the point: handed a moodboard and a request, an image model tends to reproduce the moodboard's subjects instead of its style.
Rebuild a scene as a layered paper pop-up: cut shapes, visible depth, a shallow book gutter. The cut-paper look is the constraint — flatten the layers and it reads as an ordinary illustration.
Write a frame the way a cinematographer would — lens, aperture, height, angle, movement — and render it. The model call translates a plain description into that vocabulary, which is what actually moves an image model off its default eye-level 50mm look.
Re-imagine a product as if it were made of something else, keeping its form exactly. The silhouette is held so the result still reads as the product — a bottle carved from ice is interesting, a block of ice is not.
Four costume variants of a single character, laid out side by side with the same pose and lighting so they can be compared rather than admired. The pose is pinned in the prompt because a costume sheet where the figure also moves is a mood board.
Nine hairstyles on the same face in a 3x3 grid, with the face held constant so the comparison is about the hair. One generation, because nine separate edits change the face more than they change the hairstyle.
Stick cartoon googly eyes onto whatever is in the photo. Deliberately trivial, and useful as the smallest possible worked example of an image edit that must leave everything else alone.
Continue a scene: describe what happens next, then draw it with the setting and characters held. The written continuation exists so the second image is a consequence of the first rather than a variation on it.
Four alternatives to an image, each varying composition and treatment while keeping the subject. Four separate edits rather than a grid, because each is meant to be picked and used on its own.
A character laid out as separate limb pieces on a transparent-ready sheet, the way a cut-out animation rig is assembled. Overlap is what makes a rig sheet unusable, so the prompt forbids it explicitly.
See a brand on the things it will actually appear on — a tote, a cup, a shopfront, a van. One generation containing all four, so the identity is applied consistently rather than drifting between separate renders.
Translate the copy in an ad creative and set it back into the same layout. The translation is done by a language model first, because an image model asked to translate tends to invent plausible-looking words in the target script. Even so, check the result — the type is rendered, not typeset.
Nine film stills from one imagined scene laid out as a 3x3 grid — a look-book for a film that does not exist yet. Generated as one image so the grade, lens and wardrobe stay constant across the nine, which is the whole point of the format.
A filmic colour and contrast pass on a photo: lifted shadows, controlled highlights, a considered palette. NOT AN UPSCALE — this changes the grade, not the resolution. The dump's equivalent flow used an upscaling model; this repository has an `upscale` modality but no slug it can confirm, so promising more pixels here would be a lie in the title.
Restyle a portrait into a stylised game character while keeping the person recognisable. The likeness constraint is the whole difficulty: push the style far enough and it is somebody else, which is the failure everyone notices immediately.
Generate a tangent-space normal map from a flat texture, in the lavender-blue convention shaders expect. Worth saying plainly: this is a plausible normal map, not a measured one — it is a starting point for look development, not a scanned surface.
A single continuous white line drawing on black — the minimal poster treatment. Continuity is the constraint that makes it work and the one a model breaks first, so the prompt asks for it twice.
Place an app screenshot into a phone held in a real setting, for a store listing or a landing page. The screenshot is composited by the image model rather than by `image.composite`, because it has to be warped onto the screen at the phone's angle — a flat overlay at a gravity position would sit on top of the photo, not in it.
Fit a skeleton to a mesh you already have, so it can be animated. One step, and it takes a mesh rather than a prompt — supply a glTF and nothing else. Useful on its own and as the tail of any pipeline that produces geometry: bind `mesh_url` to a previous step's `outputRef` instead of the workflow input.
Free trial credits, no card. Start from a template above, or see how a workflow is put together on the canvas.
