Skip to content

He Stopped Writing Prompts and Let an Agent Derive Them From Two Years of His Own Work

A Japanese AI filmmaker says his own prompts and the ones ChatGPT invented for him wasted credits in exactly the same way. His fix was to hand an agent two years of his production archive and let it write the prompts. What the method is, where the claim is thin, and why he refuses to automate it.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A long shelf of labeled archive boxes in a dim room, one open box spilling index cards onto the floor
Editorial illustration; not a screenshot of the described workflow.

An independent filmmaker in Japan, who publishes as CreativeEdge, made episode two of his series COLORS with Seedance 2.0 and put it on YouTube in August. Less than four months later he says watching it makes him want to delete it. Not because the story got worse. Because the footage already looks like it came from another era of the tool.

Episode three is 58 seconds of the same story, built on Seedance 2.5, and he says it finally does the things he gave up on in the first two episodes. The interesting part is not the model upgrade. It is what he changed about who writes the prompts.

He traced his wasted credits to two habits that feel different and are not: writing prompts himself, and asking ChatGPT or Claude to invent prompts for him. Both produced a lot of failed generations. Both burned the same credits on the same blind guessing. What replaced them was an agent that develops prompts from two years of his own production archive, and his argument for why that works is the most useful thing in the post for anyone paying per clip.

Four months is enough to make your own work look old

The timeline is the argument. Seedance 2.0 in August, Seedance 2.5 in November, and the earlier episode reads as dated to its own author. Aesthetic drift in generative video is now faster than most people's project cycles, which means a workflow pinned to one model's quirks has a shelf life measured in months.

That has a consequence people underrate. Skill you build by memorizing syntax and settings is depreciating skill. If the tool changes under you twice a year, the durable asset is not your prompt library. It is the record of your own decisions: what you tried, what you rejected, and what you were reaching for when you rejected it. An archive of outcomes survives a model swap. A folder of clever prompts mostly does not.

Two habits that cost the same credits

His diagnosis is blunt: writing the prompt yourself and asking an LLM to write it for you are the same mistake wearing different clothes. In both cases you start from nothing but a description of the shot, then hope the model converges on your intent, then generate again when it does not. Every one of those loops is billed.

The failure mode differs a little. When you write it yourself, you spend an hour trying to translate a feeling into adjectives, and the adjectives are usually borrowed. When an LLM writes it, you get a competent, statistically average prompt, which is a prompt shaped by everyone's taste rather than yours. That is why the second version often reads better and still misses: it is fluent about the thing you asked for and ignorant about the thing you meant.

His fix is to stop producing prompts from a blank page at all. He gives an agent the raw material of his own work and has it develop the prompts from that. In his account, the difference between asking a chat model to invent a prompt and having an agent develop one comes down to the depth and scale of the data you supply, not the wording of your request.

What "giving it your archive" actually means

The archive he describes is not a mood board. He says it goes back to 2014, a year in which he ran about sixty livestreams, and includes recorded video, slide decks, idea notes and test clips from roughly the following two years. Handed to an agent, that pile is asked to surface his preferences, his fixations, his biases and the areas he simply does not care about, and then to hold that profile fixed as the reference point for prompt development. He calls the result fixing himself as the absolute creator.

That phrase is doing real work. A prompt is a set of decisions. Most people make those decisions fresh each time, under time pressure, in a text box. He is trying to make them once, from a large body of evidence, and reuse them. The practical shape is a pipeline rather than a trick:

  1. A body of finished work plus its rejected alternatives, each with a note on why it was kept or dropped.
  2. A distilled profile of the decisions that repeat: framing habits, how much motion he tolerates, what he refuses to show, how a scene should end.
  3. Per-shot prompts derived from that profile and the scene's dramatic purpose, not from a fresh description of what the picture should look like.

The order matters. Prompt is the last step, and the least interesting one. His stated reason for preferring this over his own instinct is cost, and it is a fair reason. Trial and error is the expensive part of generative video, and his claim is that the archive cuts the number of trials by making the first guess less generic. If you are looking at the actual maths of what a usable second costs, the biggest lever is usually not the model's price per clip but how many clips you throw away to get one you keep.

He pairs it with a second change he says came out of an unrelated project: fixing each character's personality, not just their face, and letting that drive performance. He reports that this noticeably raised the level of ordinary, low-drama acting in a scene. Read that as a practitioner's report, not a controlled result. It is also the part most worth testing yourself, because it is cheap: a character card with behaviour, not just appearance, costs a paragraph.

The part that is a model feature, not a technique

He is explicit that this method does not port. In his tests, Seedance 2.5 is the only model that responds to scenario-level prompting the way the workflow needs, and he points to its limit of 30 reference images as something no other model offers. For work where that specific dramatic control is not the point, he names Wan 3.0 and Kling 4.0 as substitutes, and says his AI idol business runs entirely on Wan 3.0.

That admission is what makes the rest credible, and it also deflates the easy version of the lesson. You cannot copy "let an agent write my prompts" and expect the same result on a different model, because the method leans on capabilities that are rented, not owned. What transfers is the structure: an explicit record of your taste, a character definition that includes behaviour, and prompts derived from both. Those survive a model change. The 30-reference ceiling does not.

Where the claim is thinner than it sounds

There is one practitioner here, with one project, no A/B test, no count of attempts per shot and no credit receipts. "It works better" means "I stopped wasting credits", which is his judgement and may also be explained by four months of practice and a newer model. Anyone quoting this as proof that agent-developed prompts beat human-written ones is overreading a blog post.

Two more limits are worth naming because they apply even if his method works exactly as described.

An archive anchors you to who you were. Two years of 2014-era taste is an excellent instrument for consistency and a poor one for exploring a look you have never made. Use it when the job is to hold a line, and expect it to pull against you when the job is to break one.

An archive without failures teaches the wrong lesson. If you only feed the model the work you shipped, it learns to reproduce the hits and not the judgement that separated a hit from the near-miss three takes away. The rejections and the reason for them carry more information than the successes.

There is a rights question too, and it is not hypothetical for anyone with client work. Feeds assembled from recorded sessions and past projects contain other people's faces and material you may not have licensed for model input. Deciding to mine your own archive is a legal decision before it is a workflow decision.

Why he refuses to automate it

The most contrarian line in the post is about tooling. He argues against wiring generative pipelines into rigid automation, because these workflows are rebuilt so fast that a tightly coupled system becomes a liability. He keeps his own deliberately loose, so a model can be swapped, a node deleted, a step replaced without a rewrite. His reasoning: the current best practice may be unrecognisable in six months, so formalising it heavily is a bet against your own future.

That is the correct instinct, and it is the opposite of how most people respond to a workflow that finally works. The moment something produces good output, the temptation is to freeze it. But a frozen pipeline bakes in the model's current behaviour as if it were a law of the medium.

A useful split is stable layer against disposable layer. The taste profile, the character definitions and the scene bible are stable; they are about your work. The prompt syntax, the reference limits, the node graph and the model choice are disposable; they are about the vendor's current build. Put the two in separate places, and a model swap costs an afternoon instead of a rebuild. The same logic is why a prompt worth keeping is closer to a skill than a line of text: the durable part is the context and the rules, not the wording.

What to start recording this week

You cannot derive prompts from an archive you never built, and the archive has to predate the moment you need it. Start with what you already have and add a ledger from today.

For each generation, keep four things: what you asked for, what came back, whether you kept it, and one sentence on why. Export the keeps at full resolution and keep the near-misses too. Every couple of weeks, read the reasons and write down the patterns that repeat. That document, not the folder of prompts, is the thing that will still be true after the next model release.

Then run the comparison on one scene. Write a prompt the way you normally do, generate, and note the credits spent to reach something you would ship. Derive a prompt for the same shot from your profile, generate, note the same. Then count the hours the archive and the profile took to build, and the agent time and token cost of deriving the prompt, because credits are only half of the cost and the comparison is not honest without them. Two data points on one scene are an estimate, not a study, but they are yours, and they will tell you faster than anyone's blog post whether the archive is one you will earn back. If your work is short-form video where a usable second is the unit of spend, that comparison is the whole business case.

The rest of this workflow is about references and how they hold a character across shots, which we cover in the Seedance 2.5 multi-reference guide. What the Japanese post adds is the step before that: deciding, once and in writing, who is doing the choosing.

Sources

Share this guide
X LinkedIn