Ritish

Building LLM agents + AI tools in public. HydPy × NVIDIA Nemotron runner-up (CodeMigrator). AI workflows & sites for small businesses → Rynavo. Hyderabad

[Ritish @KurmaRitish]

A 14-second fly-over of PromptReel.

AI motion videos with the exact prompt behind each one, agent skills, trending repos and free website sections. All free.

https://t.co/ZSPhsB9IhK

Music: "Inspired" by Kevin MacLeod, CC BY 4.0

[Ritish @KurmaRitish]

@0xSilver_Time Most of those lists are general text prompts. For video, a prompt is only useful next to the clip it made, so I built a gallery where every motion video sits beside its exact prompt, credited to the creator. It's free: https://t.co/0YUT8H0hzz

[Ritish @KurmaRitish]

@NarasiVisual Going from CapCut Reels to a Claude + Remotion edit with no editing background is a big jump. Keep the prompts that worked as reusable pieces: I built a gallery that shows each video next to its exact prompt or skill, credited: https://t.co/0YUT8H0hzz

[Ritish @KurmaRitish]

@lgs1520 Code animation wins on exactness: the same input gives the same frame every time, which video models can't promise. I keep a gallery of prompts and skills from people doing this, credited to them: https://t.co/0YUT8H0Pp7

[Ritish @KurmaRitish]

@kloufrenx Timing is where the gap shows up for me too. A tip: put the key beat's second range in the prompt, like 4-6s, and compare models on that beat only. Then you're judging one moment instead of the whole clip.

[Ritish @KurmaRitish]

@vinayakbuild @pcamp96 @claudeai Thanks for the concrete breakdown - the 2-core screenshot fallback and the /tmp disk gate explain the 7x realtime well. Did you find that forcing a concurrency flag kept it in the faster path, or was /tmp the hard wall regardless?

[Ritish @KurmaRitish]

@Spinzu Right - the opening beats set scope, tone and the constraint, and most failures come from copying the style while missing the constraint. When you hit a good one, do you reverse-engineer the structure and rewrite, or keep the original phrasing?

[Ritish @KurmaRitish]

@WhiteAlchemyLab 5 cut scenes with explicit cam moves is a solid way to keep the model honest. For intense emotions, lock the same reference image across cuts - the prompt steers the emotion, the ref locks the face. Did locking the ref hold the turns, or did the emotion bleed between shots?

[Ritish @KurmaRitish]

@XunDavid80270 The rule of capping re-rolls at 3 and then changing the storyboard instead is the best part. Most people burn credits re-rolling a weak shot list. A fixed character card copied word for word each version also helps a lot.

[Ritish @KurmaRitish]

@Manojkumar2257 Captions first, for me. They are the most repetitive part and the easiest to check by eye. I keep a page of trending video and motion skill repos, with remotion-dev/skills in the mix, so it's easy to compare options: https://t.co/MdhlvqRlvr

[Ritish @KurmaRitish]

@andrewolinek Giving the model a director persona plus a scene-by-scene structure is what keeps music videos coherent. I'd split it so each scene prompt restates the visual style in one line, because style drift between clips is where multi-scene videos usually fall apart.

[Ritish @KurmaRitish]

@VibeCoderOfek @0xSubarru Treating a broken transition as a source edit is the real win of Remotion over generative video. I build with the same idea, so I track repos that pair Claude Code with Remotion here: https://t.co/MdhlvqQNFT

[Ritish @KurmaRitish]

@JamesWilsovi51 @higgsfield_ai Render time beating prompt time is a good problem to have. I find the slow part becomes choosing wording, so I keep working prompts with their results side by side to start from: https://t.co/0YUT8H0hzz

[Ritish @KurmaRitish]

@StunLabs Describing how elements connect, not just what they are, is exactly what stops a prompt from producing a pile of objects. The harbor-in-a-bottle example makes that clear. Does the same trick hold for motion, like camera path through the bottle?

[Ritish @KurmaRitish]

@SadiaMalik182 Splitting it into six timed scenes with audio and format spelled out is why this works as a Seedance prompt. I saved it with your name on it in a gallery I built, next to the model used: https://t.co/0YUT8H0Pp7

[Ritish @KurmaRitish]

@le0xet 700 videos from one loop makes the Prompt step the one that decides quality. Do you template it per source video, or let Opus write each prompt from scratch? Curious how much drift you see after a few hundred runs.

[Ritish @KurmaRitish]

@will123195 @TriadBrad @petergyang @jake11moran @HyperFrames_ @higgsfield Useful number, since most Seedance posts skip cost. I'd be curious how many of the $100 of generations were retries versus keepers. I keep each prompt next to the model that made it, so reruns are cheaper: https://t.co/0YUT8H0hzz

[Ritish @KurmaRitish]

@builtbysketch Checking what sits under a tool before paying for it is a habit worth keeping, especially with video models where the wrapper markup is rarely visible. Did you compare the same prompt on both, or just the pricing?

[Ritish @KurmaRitish]

@bdc58_ Splitting the agent roles into brief writer and designer mirrors how a real studio works, and it keeps the output on-brand. Does your setup hand the brief over as a markdown file, or as a plain chat message?

[Ritish @KurmaRitish]

@motionstashco @bot One-prompt motion graphics already works for simple explainers, but exact type timing and brand colors still need a second pass. I collect prompts that get close, shown beside the result, in a gallery I built: https://t.co/0YUT8H0Pp7

[Ritish @KurmaRitish]

@FastMoroAI Holding a pixel look for a full 10 seconds is the hard part, so seeing both models pass is useful. Did you run each one more than once? A single seed can flatter either model, and the chimney drift might be luck.