What the feather button actually does, and how LUVI's prompting guides are built
The feather button next to the prompt runs a per-family manual over what you wrote, for free. It has two layers: eight deterministic checks that only fire on things the manual can prove, and a dialect rewrite for the judgment. Here is what each layer can and cannot do.
Every model family reads prompts differently, and the differences are not small: a quotation mark means speech on one family and on-screen text on another, an aspect ratio typed into the prose does nothing on most of them, and a word like "masterpiece" is inert on some and actively degrading on others. The feather button beside the prompt box exists because nobody should have to hold eighteen dialects in their head. It applies the manual for whichever model you have selected, and it costs nothing to run.
What happens when you press it?
Two layers run, in order, and they have very different jobs.
The first is deterministic: a set of checks that only fire on things the manual can prove about a string. They either rewrite a token, or warn, and each one cites the section of the manual it came from. The second is the dialect: the family's prose rules, which handle everything that needs judgment—structure, emphasis, what to describe and in which order.
The split is deliberate. If a rule can be stated as "this exact string is inert on this family", it belongs in the first layer, where it is exact and free of opinion. If applying it requires reading your intent, it belongs in the second.
Layer one: the eight things a manual can prove
There are eight check types in the deterministic layer, and between them they cover most of the habits people carry between models.
- references — rewriting a reference into the family's binding form. On Nano Banana,
image 1becomesImage 1, because that family binds references as ordinals. - weights — stripping Stable Diffusion weight syntax while keeping the word.
(melancholy:1.4)becomesmelancholyon every family that has no cfg. - editPreserve — appending a preservation clause to an edit instruction that lacks one, because on some families an edit regenerates the whole frame.
- deadTokens — deleting words a manual names outright. On Seedream, "masterpiece, best quality, 8K, ultra-detailed, trending" all disappear in one move.
- charCaps — warning before a prompt overruns a documented character limit.
- channels — the typed audio channels on the families that use them.
- temporal — how a family reads timestamps, which is the rule that differs most violently between otherwise similar models.
- proseWarns — everything else a manual states as a flat fact: an f-stop that does nothing here, a hex color that does nothing there, a resolution that belongs in a parameter.
A single pass can fire several at once. A prompt carrying a ratio, a resolution, a film stock, an f-stop, a weight, a hex color, a duration and a request for silence came back from the FLUX 3 Video check with one rewrite and eight warnings—each one naming the section of the manual behind it.
Layer two: the dialect
The second layer is the family's own prose manual: its formula, its vocabulary, the order it rewards, the things it silently ignores. Eighteen model families have one.
It is assembled as a fixed block of text that is identical byte for byte for a given family and mode. That sameness is not an aesthetic choice—it is what keeps the check free to run as often as you like, which in turn is what makes it worth pressing on every prompt rather than only on the ones you are worried about.
Why do some models have no guide?
Because not every vendor manual contains a rule that can be enforced. Two families on LUVI deliberately have no deterministic profile at all: Veo and FLUX 1.1 Pro. Their manuals contain no unconditional formal string rule—Veo's timestamps are valid syntax rather than a mistake, and the FLUX 1.1 advice that looks like a rule turns out to depend on which mode you are in. Both still have a dialect. They just have nothing the first layer could honestly assert.
This is why the check sometimes reports that a model is dialect-only. It is not a gap in coverage; it is the manual being accurate about what it knows.
What will it not do for you?
The judgment calls, and they are the expensive ones.
It will not turn "no cars" into "an empty, deserted street"—that rewrite requires knowing what you meant, and a wrong guess is worse than a warning. It will not reduce five stacked camera instructions to the one the family can actually execute. It will not decide whether your audio line needs a music bed. Those belong to you, and the manual's job is to tell you they exist.
It also cannot set a parameter. When a ratio in your prose gets flagged, the fix is the aspect or size control beside the prompt, and the check has no hands to reach it.
How we checked
- What was read: the deterministic check implementations and every registered profile and dialect in the prompt engine, plus live runs of the check against prompts on FLUX 3 Video, HappyHorse, Nano Banana Pro and Seedream 5.0 Pro.
- Date: September 22, 2026.
- Results: eight deterministic check types; thirteen families carrying a deterministic profile; eighteen families carrying a dialect; two families (Veo, FLUX 1.1 Pro) documented as dialect-only by design.
- Drawbacks: No generations were run for this post beyond the free prompt checks, so it makes no claim about how much the rewrites improve an output—only about what the check does to the text.
In LUVI
The button is the feather beside the prompt box in the Workspace, and the help center describes it under what the feather button does and what a prompting guide is. Checks and rewrites do not cost credits.
If you work through Claude or ChatGPT, the LUVI connector exposes the same two things as tools: one that hands the assistant the family's manual before it writes anything, and one that runs the check on a finished prompt. That is the better order when the assistant is drafting for you—give it the manual first, and the check has less to correct.
New here? Create a LUVI account, paste in a prompt you wrote for some other tool, and press the feather before you spend anything. The list of things that would have done nothing is usually longer than people expect.
Sources
- How to create effective prompts with Veo 3 — Google DeepMind, prompt guide. Accessed September 22, 2026.
- FLUX prompting guide — Black Forest Labs, documentation. Accessed September 22, 2026.
More from Inside LUVI
- What is LUVI Creator? How one workspace for AI video, image, audio and 3D makes production faster and betterLUVI Creator is a web platform for generating video, images, audio and 3D with 223 AI models from 30 families, in one account paid in credits. It speeds work up with parallel runs, reusable references and workflows, and protects quality with prompting guides for 18 model families.
- Estimate vs charge: why the number you pay only ever goes downLUVI shows a credit estimate before every generation and charges you when the result arrives. Since September the charge follows what the job really used — capped at the estimate, never above it. What changed, what we measured, and why one model dropped from 15 credits to 7.
- Why every image you upload to LUVI becomes JPEG or PNGiPhone HEIC photos used to fail in Chrome, WebP confused half the models, and one library's browser decoder took 27 seconds on a 12-megapixel shot. So we moved the conversion to the server and made one rule: every image is stored as JPEG or PNG, colour profile and orientation intact.