Writing license terms for prompt packs and LoRAs
Four decisions decide whether a buyer comes back: commercial use, redistribution, retraining, and derivative outputs. Written out as plain-language tiers.
The question that kills a repeat sale is not "how much". It is "can I use this for a client?", asked in a DM, three days after purchase, because the licence file said "for personal and commercial use" and nothing else.
That phrase is worse than no licence at all. It answers one question and implies answers to three others it never addresses, and every one of those gaps eventually surfaces as a support thread, a chargeback, or a buyer who quietly never returns. A licence that a buyer can read in ninety seconds and act on without asking you anything is a retention feature, and it costs one afternoon to write.
There are exactly four decisions. Make each one explicitly, in a sentence a non-lawyer can act on.
Decision 1 — Commercial use
Not "yes or no". Three sub-questions, and buyers hit all three.
Can the buyer sell what they make with it? Prints, merch, ads, client deliverables. If the answer is yes, say which of those, because "commercial" means different things to a print-on-demand seller and an agency.
Can they use it in client work? This is the one people assume and shouldn't. A freelancer delivering a campaign built on your LoRA is a different exposure from someone posting to their own feed, because the client will ask who owns what.
Is there a revenue or seat ceiling? Some sellers cap commercial use at a revenue threshold or a team size and sell an extended tier above it. Legitimate, but only if the number is stated. An unstated ceiling is not a ceiling, it is a surprise.
Plain-language version that works:
Commercial use: Yes. You may sell images, video, merchandise and client deliverables made with this pack. There is no revenue cap and no per-project fee. Client work is included — your client may use the finished assets they commissioned, but they do not receive a licence to the pack itself.
That last clause is the one most licences omit and most disputes turn on.
Decision 2 — Redistribution
Redistribution is what protects your revenue, and it is the decision where vague wording does the most damage. Break it into three cases, because they are genuinely different.
| Case | What it means | Common answer |
|---|---|---|
| Sharing the files | Buyer uploads the pack or the LoRA weights somewhere others can download | Prohibited |
| Sharing inside a team | Buyer's colleagues use it on the same projects | Depends on your tier — say so |
| Bundling into their own product | Buyer folds your prompts into a pack they sell | Prohibited by default, licensable separately |
The third case is the one that catches people out. A buyer building their own workflow pack does not think of including your prompts as redistribution — they think of it as using what they bought. If you want that stopped, say the word "bundle" explicitly. "You may not redistribute" does not obviously cover it in a buyer's head.
The team case deserves a real number rather than a shrug. "One seat" and "up to five people at one company" are both fine positions. "Personal use" is not, because a buyer with two colleagues cannot tell which side of it they are on.
Decision 3 — Retraining
This one applies to model artefacts specifically, and it is where an unstated position costs you the most. If you sell a LoRA or any fine-tune, the buyer holds weights. Weights can be merged into other models, used to generate a synthetic dataset, and used to train a replacement that captures your style without containing a byte of your file.
Three positions, all defensible, but you have to pick one:
- No retraining or merging. The strictest. Simplest to state, hardest to detect a breach of.
- Merging allowed for personal use, not for redistribution. The most common practical middle ground — buyers get to experiment, the merged result cannot be published or sold.
- Retraining allowed with attribution. Suits creators whose goal is reach rather than exclusivity, and it pairs naturally with an open-distribution strategy.
Write it in terms a buyer recognises rather than in the abstract:
Retraining and merging: You may merge this LoRA with other models for your own use. You may not publish, share or sell a merged model, a derivative LoRA, or any model trained on outputs generated with this one.
The phrase "trained on outputs generated with this one" closes the synthetic-dataset route. Without it, the restriction has a hole large enough to drive the whole use case through. Enforceability across jurisdictions is genuinely unsettled, but a stated position gives you standing to act on a platform report and an unstated one gives you nothing.
Still deciding whether to ship a trained artefact at all? Style references versus fine-tunes covers which one actually locks a look, and what a trending LoRA list tells you covers where that market is moving.
Decision 4 — Derivative outputs
Who owns the image the buyer generated? Sellers reflexively answer "they do", and then write a licence that quietly contradicts it.
Two things to get right.
Say the buyer owns their outputs, subject to the model's own terms. You are licensing your pack, not the upstream generator. The generator's terms of service apply on top of yours, and buyers do not always realise that. One sentence saves the confusion.
Do not claim a share of their revenue. Some licences reserve a royalty on outputs. It is very hard to police, it makes commercial buyers walk, and it turns a $29 purchase into a legal review. If you want recurring revenue, sell a subscription; do not tax outputs.
The copyright position underneath this is worth knowing before you write the clause. The US Copyright Office has held that a purely AI-generated image with no meaningful human creative input is not copyrightable, so a raw prompt output has no protection your licence can lean on. What creates a protectable work is human contribution: arrangement, composition edits, typography, selection and sequencing. That cuts both ways for your buyer — they own their outputs in the sense that you are not claiming them, and may still have nothing enforceable against a copyist unless they did real work on top. Saying so plainly in the licence is unusual and buyers remember it. The legal and licensing overview for AI content goes deeper.
Two tiers, written out
Most sellers need exactly two. Put them side by side on the sales page so the buyer self-selects instead of asking.
| Personal | Commercial | |
|---|---|---|
| Use in your own projects | Yes | Yes |
| Sell outputs (prints, merch, ads) | No | Yes |
| Use in paid client work | No | Yes |
| Team seats | 1 | State a number |
| Share the files | No | No |
| Bundle into a product you sell | No | No |
| Merge / retrain | Personal experiments only | Personal experiments only |
| Publish a merged or derivative model | No | No |
| Ownership of outputs | Buyer, subject to model terms | Buyer, subject to model terms |
| Attribution required | No | No |
On the last row: requiring attribution on commercial work sounds harmless and is a real friction point. Agencies cannot always credit a tool in a client deliverable, and a licence that demands it loses those buyers. Make attribution a request, not a condition.
On the first column: a personal tier that permits nothing commercial only makes sense if you police the difference, and most sellers do not. If you are not going to, sell one tier that includes commercial use and price it accordingly. A single clear licence beats two blurry ones.
Put it in three places
The licence only prevents support threads if the buyer meets it before they need it.
- On the sales page, as the two-column table above. Not a link to a PDF. Buyers decide here.
- In the download, as
LICENSE.txtat the root of the zip, in the same plain language. - In the changelog, if the terms change. Say which version the change applies from, and honour the old terms for people who bought under them. Retroactive tightening is the fastest way to lose a repeat buyer.
Two terms buyers will bring to you: usage rights and exclusivity windows, both of which surface the moment an agency is on the other end.
FAQ
Do I need a lawyer to write this?
For a $29 pack sold to individuals, a clear plain-language document you wrote yourself is far better than a copied boilerplate you do not understand, and better than nothing by a wide margin. Get advice before you sell enterprise or exclusive tiers, before you write revenue caps you intend to enforce, and before you build a business where the licence is the product. Nothing here is legal advice.
Can I actually enforce a no-retraining clause?
Enforcement in the general case is unproven and jurisdiction-dependent. What a stated clause reliably gives you is standing on the platform where the breach appears — a marketplace or model host will act on a takedown request that points at explicit terms far more readily than one that points at an implication. Treat the clause as the thing that makes a report actionable rather than as something you will litigate.
Should my LoRA licence match the base model's licence?
It has to be compatible with it, which is a different and stricter requirement. If the base model's terms restrict commercial use or downstream redistribution, you cannot grant your buyers more than you hold. Read the base model's licence at its source before you write yours, and if your terms are more permissive than the base allows, yours are the ones that are wrong. Some open weights also carry territory or field-of-use restrictions that a buyer will not think to check, so name the base model in your own licence rather than leaving them to work it out.
What changes if I sell to agencies rather than individuals?
Three things. Seat counts stop being theoretical and need a real number. Client-deliverable rights become the first thing procurement asks about, so put that clause near the top. And you will be asked for an exclusivity option — decide in advance whether you sell one and at what multiple, because being asked cold is how people give away a category for a small fee. The packaging side is covered in digital product sellers.