17:14
AI video generation
The video demonstrates a three-tool workflow — a custom Claude skill that breaks a short story into timestamped, segmented prompts, Higgsfield's Seedance 2.0 for generating the actual video clips from those prompts and reference images, and Gemini's free "Past Forward" app for producing decade-styled reference photos of the creator — that let him produce a roughly 30-second time-travel short film in under an hour, after an early attempt at single long generations failed and he switched to shorter, cheaper 8-second segments.
Three-tool workflow: Claude (skill) for prompt/script generation, Higgsfield's Seedance 2.0 for video generation, Gemini's "Past Forward" app for decade-style reference images.
The Claude skill takes a short story (typed or spoken via text-to-voice) and breaks it into scene-by-scene prompts with timestamps, embedding the goal of the video into the instructions.
Higgsfield's Seedance 2.0 interface: upload reference media/photos/videos/sounds, turn uploads into tagged "elements," write a prompt, and set model, generation length, aspect ratio (16:9), and resolution (720p or 1080p).
Cost/resolution tradeoff: an 8-second 720p generation costs 32 credits versus 80 credits for 1080p; a single 15-second generation cost 150 credits.
Seedance 2.0's generation limit is 15 seconds; the creator's first attempt split an 18-shot/33-second script into two roughly-15-second halves, producing a disjointed, failed result that didn't loop back to the intended ending.
He revised the Claude skill to output 8-second segment prompts instead of one long prompt, which gave more control and cost less per attempt, and became the final production method.
Reference images from Gemini's "Past Forward" (decade-styled selfies spanning the 50s through 2000s) were uploaded to Higgsfield and tagged into prompts (e.g., "the bearded figure @image1") to anchor the character's appearance per decade.
A previously generated video clip was itself fed in as a reference for a later generation, specifically to visually close the narrative loop back to the opening scene.
Iterative prompt debugging: when a generation showed a visual error (a watch facing backwards), the creator described the problem back to Claude, which revised the prompt; the error persisted across two regeneration attempts before a workable (if imperfect) result appeared on the third try.
Post-production cropping was used to hide a generation artifact (an extra third hand appearing near the watch) rather than spending more credits on another regeneration.
Generation order in production was non-linear: the 1990s scene was generated first, followed by the intro/loop-closing 1960s scene, based on which segments could be salvaged from earlier failed 15-second attempts.
Total time breakdown: about 15 minutes to build/refine the Claude skill, under 2 minutes for the Gemini Past Forward image generations, about 20 minutes for the Higgsfield video generations, and about 20 minutes of editing — a little under an hour overall.
Seedance 2.0 (via Higgsfield) — An AI video generation model accessible through the Higgsfield platform, used to generate video clips from text prompts and uploaded reference images/elements. Apply: Upload reference media as tagged elements, write a scene prompt, set model/duration/aspect ratio/resolution (720p or 1080p), and generate the clip.
Claude skill for script segmentation — A custom Claude skill that takes a short narrative and breaks it into scene-by-scene prompts with timestamps for feeding into a video generator. Apply: Write or speak a short story into the skill's chat, then copy each segment's prompt output directly into the Higgsfield prompt box for that scene.
Gemini "Past Forward" (Google AI Studio) — A free Gemini app-studio tool that takes an uploaded photo of a person and generates decade-styled versions of them, from the 50s through the 2000s. Apply: Log in with a Gmail account, upload a selfie, click generate, then download the decade-specific images to use as reference images in Higgsfield.
Element tagging / reference-image invocation — Higgsfield's mechanism for referencing an uploaded image inside a prompt by tagging it (e.g., "@image1") so the model anchors the character's appearance to that image. Apply: Upload the relevant decade photo, then reference it at the point in the prompt describing the character, e.g. "the bearded figure @image1".
8-second segmented generation (vs. 15-second limit) — A production strategy of generating video in 8-second chunks instead of Seedance's roughly 15-second single-generation limit, to gain more control and lower cost per attempt. Apply: Revise the Claude skill to output prompts in 8-second segments per scene, then generate and review each segment individually instead of submitting one long multi-shot prompt.
Video-as-reference chaining — Feeding a previously generated video clip, not just a still image, into a new generation's reference input to maintain visual continuity, e.g. to close a narrative loop. Apply: When generating a final scene that must visually match an earlier one, upload both the anchor reference image and the earlier generated video clip as references for the new prompt.
Iterative prompt debugging with the LLM — A feedback loop where a flawed generation (e.g. a backwards watch) is described back to Claude, which revises the prompt in an attempt to fix the visual error. Apply: Identify the specific visual defect in a generation, describe it to the LLM that wrote the prompt, and resubmit the revised prompt to Seedance, repeating if the defect recurs.
Post-production cropping to hide generation defects — An editing technique of cropping/cutting a clip to remove a visible AI generation artifact, such as an extra third hand, rather than spending more credits on regeneration. Apply: Crop the frame tightly to exclude the defective region for the portion of the clip where it appears, then cut to the next usable segment.
Resolution/credit tradeoff (720p vs 1080p) — Higgsfield's Seedance 2.0 pricing distinction where 1080p generations cost significantly more credits than 720p for the same clip length. Apply: Default to 720p (e.g. 32 credits for an 8-second generation) rather than 1080p (80 credits) to conserve credits across multiple generation attempts, reserving 1080p for when higher quality is specifically needed.
The stated bottleneck wasn't the video model itself but prompt segmentation strategy: switching from one long 33-second script to discrete 8-second segment prompts is described as the single biggest time investment and the fix that made the workflow actually work.
The effective cost of a finished short is higher than the advertised per-clip credit price: the first 15-second/150-credit generation was a total loss, and later on "300-plus credits" were wasted on two more discarded generations before a usable watch-scene clip was produced.
Video-as-reference chaining (feeding a finished output clip, not just a still photo, into the next generation's reference slot) is used deliberately to preserve visual continuity when a scene needs to match an earlier one exactly, such as closing a story loop.
Prompt revision doesn't reliably fix specific visual defects: the same backwards-watch error survived a revised prompt and reappeared on a second attempt; the third attempt introduced a new defect (an extra third hand) that was ultimately fixed by cropping in post rather than by further regeneration.
Production order followed opportunistic salvage logic rather than the story's chronological order — the creator generated scenes based on what could be reused from earlier failed 15-second attempts, not in narrative sequence.
«All right, this whole thing from idea to finish to the editing took about an hour to make and that is possible because of three different tools that I'm going to show you in this workflow.»
— 00:45
«So it's 80 credits versus the 32 credits if you do a 720 generation.»
— 02:08
«The first tool is Claude. So Claude skill is what I'm using to generate these prompts.»
— 02:43
«But the best thing about having a skill or a custom GPT or a specific set of directions is that it understands what the goal is.»
— 05:51
«I just wanted to show you this failure because it's just part of creating, right?»
— 08:06
«So I went back to my skill and instead of it giving me all the different prompts for the entire thing, I wanted to break it down with 8 second generations shot by shot»
— 08:59
«This is where he's like, "Oh man, not this again."»
— 12:46
«I ended up scrapping both of these generations. So, it's 300-plus credits wasted.»
— 13:19
«This is the trial and error that it takes for you to create something like this.»
— 15:11
«the biggest time consumer was really revising the skill so I can go from these long 15-second generations to eight-second bits.»
— 15:38
Reception
Engaged audience with genuine interest in the tools, but tempered by practical concerns about cost, authenticity, and whether alternatives might be easier.
The video functions as a practical build-log tutorial mapping a specific, reproducible three-tool pipeline (Claude skill → Higgsfield Seedance 2.0 → Gemini Past Forward), and is unusually transparent about failed generations, persistent model errors, and wasted credits, making it a useful reference for the real iteration and cost involved rather than a polished demo.
17:14