Turning vague client notes into changes
A translation table from client language to testable production changes, plus the three questions that resolve a vague note before you spend attempts.
"Make it feel more premium" has at least six causes and no single fix. It might be the lighting, the amount of stuff in frame, the pace of the cut, the lens, the grade, or the typeface on the end card. Each of those is a different edit, most of them are cheap, and picking the wrong one costs you a round — not because of the credits, but because you send back a version that changed something the client did not mean and they quietly conclude you were not listening.
Translation is a skill with a method. It has three parts: a table you can look things up in, a small set of questions that resolve the genuinely ambiguous notes, and a discipline of changing one variable at a time so you learn something from every attempt.
Why guessing is more expensive than asking
The attempt budget is real even if nobody has measured it properly. Nothing usable exists in the public record on how many generations it takes to land a shot, and the figures that circulate are aggregated anecdote. Your own number is the only one worth planning against, and once you have it the logic is unavoidable: attempts are the scarce resource on a deadline, and a guessed interpretation spends several of them proving a hypothesis nobody held.
Asking a clarifying question costs one message and about four minutes of client time. It is almost always the cheaper move, and it has a side effect that compounds: the answer usually becomes a reference you can reuse for the rest of the account.
The translation table
Look the note up, then confirm the reading with the client rather than acting on it silently.
| Client note | What it usually points at | Concrete change | Where it lives |
|---|---|---|---|
| "More premium" | Clutter, flat light, restless camera | Fewer props, single soft key with deep falloff, longer lens, near-static push, slower cut rhythm | Prompt + edit |
| "It looks cheap" | Frontal light, plastic surfaces, default type | Motivated side or rim light, visible material texture, simpler background, replace on-screen type in the editor | Prompt + edit |
| "More energy" | Cut rhythm, not brightness | Shorter shots, earlier hook, handheld or push-in, faster subject motion | Edit + prompt |
| "It doesn't feel like us" | Brand manual failure, not taste | Check hex values, typeface, and copy tone against the manual before changing anything creative | Brand check |
| "It feels too AI" | Artefacts and unmotivated movement | Name the specific tell — hands, on-screen text, weight and contact, over-smooth skin, drifting camera — and regenerate that shot | Regenerate |
| "The product looks wrong" | Reference fidelity | Rebuild the shot from an approved still of the real product rather than from text | Key frame |
| "Make it pop" | Hierarchy, usually in the graphics | Raise title contrast against its background, one focal point per frame, larger type | Edit |
| "More authentic" | Camera and audio, not content | Handheld framing, imperfect eyeline, room tone, minimal grade | Prompt + edit |
| "It's too long" | The setup, not the whole thing | Cut the run-up, move the hook to the first frame, keep the payoff intact | Edit |
| "Something about the ending" | The final hold | Land on the product, extend or shorten the last beat by half a second | Edit |
Three patterns are worth naming because they change how you triage.
"Doesn't feel like us" is a compliance check, not a creative note. Run it against the brand manual first. Roughly half the time the answer is a hex value or a typeface, which is a five-minute fix and not a conversation about taste. Getting exact brand values out of prompts covers holding colour through generation rather than correcting it afterwards.
"Too AI" is almost always a specific artefact. Ask which second. It is usually hands, text in frame, or something contacting a surface without weight — the failures in physics and contact in generated motion rather than a general aesthetic. Naming the tell converts an unanswerable note into a shot to redo.
Light and camera are the two levers that move "feel" the most. Lighting prompts and camera movement prompts are the vocabulary for both, and when a description is not landing, handing the model a reference clip instead of describing the camera is often faster than another adjective.
The three questions
Most ambiguous notes resolve with these, asked together in one message so the client answers once.
- "Which second stops working for you?" A timestamp converts an opinion about the whole piece into a problem with a location. It also frequently reveals that the client likes 80% of it, which is worth knowing before you rebuild anything.
- "Send me one example that feels the way you want this to feel." Anything — a competitor's ad, a photograph, a film still. This is the highest-value question in the set, because the reference joins the approved reference board and makes the same note testable for the rest of the relationship.
- "Are we fixing this version, or is the direction different from what we agreed?" Asked plainly, without edge. It is the classification question that decides whether the next hour is a refinement or a commercial conversation, and clients answer it honestly far more often than people expect.
A fourth is useful when the notes are long: "If we could only change one thing, which one?" Note lists are usually ranked in the sender's head and unranked on the page.
Change one variable per attempt
Once you have a reading, resist the urge to fix everything in a single regeneration. Bundling four changes into one attempt means that when it comes back better you do not know why, and when it comes back worse you have to unpick it.
Where the model exposes a seed, hold it fixed while you change one thing, so the difference between two outputs is attributable to the edit rather than to noise. Seeds and reproducibility covers where that holds and where it does not. Keep a record of the prompt diff rather than the whole prompt — the changed clause is the thing you will want to reuse. A prompt iteration workflow is the longer version of this loop, and negative prompts are worth reaching for only after positive description has failed, not as the first move.
A worked example
The note: "Make it feel more premium."
The questions come back with: the middle section, around six seconds, when the hands come in; a still from a watch campaign as the reference; and yes, we are fixing this version.
That answer set is a diagnosis. The watch still is dark, sparse, and lit from one side; the current shot is bright, busy and frontally lit, with a camera that drifts because nothing told it not to.
Before:
woman applying serum in a bright bathroom, morning light, cinematic
After:
close-up of a hand placing a glass serum bottle on dark stone,
single soft key from camera left with deep falloff into shadow,
background unlit and out of focus, slow push in on a long lens,
no other objects in frame
Three changes, three attempts, in this order: light first (it moves the most), then clutter, then camera. Plus two edit-side changes that cost nothing to try — hold the bottle shot half a second longer, and drop the title down two sizes so the frame is not competing with itself.
Then reply naming which change addressed which part of the note: "Relit it from one side, cleared the counter, and slowed the camera — the reference you sent was doing all three." That sentence is what makes the next note more precise, because it teaches the client which words map to which levers.
When translation fails twice
If two rounds of careful translation still miss, stop translating. The note is not about the asset; it is a disagreement about the direction, and no amount of prompt precision resolves it. Go back to the reference set, put three options in front of the client, and get one of them signed. A diagnostic tree for generations that come back wrong is the checklist for ruling out the production causes first, so you only escalate when the production side really is clean.
FAQ
Isn't asking clarifying questions going to make me look unsure?
Asked vaguely, yes. Asked specifically, the opposite: "which second" and "send me one example" read as precision, not doubt. What reads as unsure is sending back three speculative versions and asking which one they prefer, because that is visibly a guess.
What if the client says "you're the expert, you decide"?
Take the authority and be explicit about the trade you are making. "Then I'll take it darker and quieter — that's what the reference you liked is doing. If that lands wrong, the alternative is brighter and faster and we can try that next." You have made a decision and pre-named the fallback, which is what they were actually asking for.
Should I show the client the prompt changes?
Show the effect, not the syntax. "Relit from one side, cleared the counter" is useful to them; the prompt text is not, and putting it in front of a client invites feedback on the prompt rather than the picture. Keep the diffs in your own working notes where they build into a reusable library.
How do I stop the same vague note recurring on the next project?
Promote the resolution into the brief. Once "premium" has been pinned to a specific reference and a specific lighting setup for this client, that belongs in the reference board at the concept stage of every subsequent job — which is where a vague note goes to stop being vague.