A prompt is not an asset

The marketplace problem

There is a small industry selling prompts. Packs of a hundred prompts for a fee, prompt marketplaces, prompt libraries as a subscription. It is worth understanding why almost all of it is worthless, because the reasoning also tells you what to invest your own effort in.

A prompt only means something in combination with a specific model, at a specific version, with specific sampler and guidance settings. Move any of those and the output changes, often completely. The models change every few months. A prompt pack is therefore a set of instructions for a machine that will not exist next year, sold without the machine.

There is a second problem. The good prompts in any pack are the obvious ones — describe the subject, name the register, state the light — and you can derive them from the last five lessons in an afternoon. The rest is the quality-incantation vocabulary that stopped working two model generations ago.

What does transfer

Some things survive a model change, and these are worth building deliberately.

Reference images. A depth map, a pose, a colour reference, a sketch. These are inputs to a process, not instructions to a particular model, and they work on whatever comes next.

Workflows. The sequence — generate at native size, inpaint the face, upscale, colour-correct in a free editor, composite the type — is a method. Node graphs in ComfyUI are files you can keep, share and adapt when the model underneath changes.

Fine-tunes you trained. A small LoRA trained on your own material encodes something the prompt cannot. It is tied to a base model, so it also expires, but it holds information no sentence can carry.

Judgement about what is wrong. This is the real asset and the reason for this whole module. Someone who can look at a failed image and say "that is guidance, not prompt" or "that is the latent resolution, no wording will fix it" is not dependent on any particular tool.

Version your own prompts anyway

Within a project, prompts are worth keeping properly. Not as a mystical library, but the way you would keep any working configuration.

A plain text file next to the project, one block per shot, each with model, seed and parameters, is enough. If you use version control for anything else, put it there; a prompt is text and diffs well. When a client asks for "the same but with a blue background six months later", the difference between a two-minute job and an afternoon of guessing is entirely whether this file exists.

The uncomfortable part

There is a related claim worth being honest about, because people building a business on prompting need to hear it: prompting is not, on its own, a durable skill. Every model release makes prompts easier to write and less necessary. Systems that rewrite your request internally have already removed much of the craft, and that direction is not reversing.

What does not get automated away is everything around the prompt. Knowing what image the job actually needs. Composition and typography. Retouching and compositing. Judging when a result is nearly right and what specifically is wrong. Managing rights and disclosure. Those are the reasons a client pays somebody, and they are all covered elsewhere in this track — design, image editing, vector work, storytelling, and the business of it.

The prompt is the cheapest part of the process and the part most likely to be obsolete. Put your effort into the parts that were already valuable before any of this existed.

That is not a discouraging conclusion. It means the ceiling on this work is set by ordinary craft rather than by access to a particular tool, and ordinary craft is learnable by anyone with time, which is the only genuinely fair distribution of ability there has ever been.

There is one narrow exception worth naming, because it is where prompt knowledge does hold value. Inside an organisation, a prompt that has been tested against a house style, checked for the failure modes that matter to that business, and recorded with its model version is a piece of operational documentation. It is valuable for the same reason a tested configuration file is valuable: somebody did the work of finding out what breaks. That value is local, it decays with the next model, and it belongs in the project's repository rather than for sale. Notice the difference from a prompt pack. What is worth something is the testing, not the sentence.