The fastest way to waste a generation budget is to open the interface before the references and decisions are ready.
I learned that the expensive way. Early on, I treated the model like a camera that was always loaded. Brief arrives, prompt goes in, takes come out, and only afterward does someone notice that the product angle was wrong, the logo version was outdated, or the motion would fight every cut the editor needed to make. The generations themselves were often fine. The preparation was not.
A thirty-second continuous take from Seedance 2.5 is long enough to carry a full commercial beat. That length is useful only if the inputs and the success criteria are locked first. Otherwise the model simply produces longer versions of the same avoidable problems.
This is the preparation checklist I now run before any controlled test. It is deliberately short. The goal is not paperwork. The goal is to make the first generation cycle informative instead of exploratory.
Why Preparation Comes Before the Prompt
Most teams I work with are already comfortable writing prompts. What they underestimate is how much of the final result is decided by the material that sits beside the prompt and the decisions that never get written down.
Seedance 2.5 can condition on a substantial set of images, short video clips, and audio. That capacity rewards clean, intentional references and punishes mixed folders of half-cleared assets. It also rewards a clear statement of what “good enough for the next handoff” actually means on this particular job.
I keep a one-page prep sheet taped inside the front of the notebook. The sheet has five sections. If any section is blank when someone is ready to hit generate, we stop.
The Five Things That Must Be Settled First
1. The Delivery Frame
Write down the exact delivery context before anything else.
Is this a social cut, a pre-roll, a website hero, an internal pilot, or a package that will later need broadcast specs? What aspect ratio and approximate length are required? Will the take need to be cut, or does it have to work as a continuous piece?
A thirty-second generation that looks strong in isolation can still be the wrong shape for the edit. Knowing the delivery frame early keeps the variation budget focused on takes that can actually travel.
I once watched a team generate a beautiful horizontal product sequence only to discover the brief had quietly shifted to vertical for the primary placement. The generations were not wasted in a technical sense. They were wasted in a schedule sense. The delivery frame should have been the first line on the sheet.
2. The Locked Elements
List every element that cannot change without triggering a regeneration or a legal problem.
Product geometry and logo version usually sit at the top. Brand colors, approved talent likeness if any, and specific claims that must remain accurate follow close behind. If a client or brand manager has already signed off on a particular angle or surface treatment, that decision belongs on the locked list.
Everything not on the locked list is fair game for variation. The distinction matters because client notes later will test it. When someone asks for “just a small change,” the locked list tells you whether the request is local or structural.
3. The Reference Package
Treat references like a camera package, not a mood board.
For a product-focused thirty-second test I usually want:
Clean white-box or controlled-lighting stills of the product from the primary angles that will appear
At least one short motion reference if the product has mechanical or liquid behavior that must stay consistent
Any approved environment or lighting stills that define the look
Audio references only if specific sound design or music timing is already locked
Quantity is less important than clarity. A smaller set of correctly labeled, rights-cleared references outperforms a large dump of mixed campaign assets. Seedance 2.5 can accept a large reference set; it cannot invent clearance or correct an outdated logo that was left in the folder.
I label every reference file with date, version, and intended use. The habit comes from years of tracing which still ended up in which composite. The same discipline prevents the model from faithfully reproducing the wrong version of a brand asset across an entire variation set.
4. The Variation Budget and Selection Criteria
Decide in advance how many takes the team will generate and how they will be judged.
A common failure mode is open-ended generation followed by exhausted review. The model can produce options quickly. Humans cannot evaluate them at the same speed without criteria.
I set a hard variation budget for any first test—usually a small number of controlled takes rather than a large exploratory set. Selection criteria are written before generation begins. Typical criteria for a product beat include:
Product remains recognizable and correctly oriented at the key moment
Motion supports the intended cut points
Lighting and surface response will survive a basic grade
No obvious artifacts that will force heavy cleanup before the next handoff
The criteria do not have to be perfect. They have to be shared. When three people look at the same take and apply different private standards, the selection meeting becomes a negotiation instead of a decision.
5. The First Revision Hypothesis
Write down one realistic client or internal note that is likely to arrive after the first review, and decide how the team will respond to it.
Examples I have used:
“Can we turn the product so the logo reads more clearly?”
“The background feels too busy against the super.”
“We need a slightly slower push on the hero moment.”
For each hypothesized note, mark whether the current process can absorb it with a local fix, a targeted redraw, or a full regeneration. That single exercise forces the team to confront the difference between a strong generation and a revisable one.
If the answer for the most likely note is “we would have to start over,” the preparation is not finished.
A Short Personal Calibration
I ran this exact checklist last week on a synthetic thirty-second test for a regional beverage brand concept. The dog was under the desk. Morgan was on a call in the other room. I had a clean white-box turn of the can, two lighting stills, and a short motion reference for the condensation behavior. Variation budget was four takes. Selection criteria were written on the same page as the locked list.
The generations came back coherent. One take clearly met the criteria. The hypothesized note—“rotate the can ten degrees at the hold”—was addressable with a local adjustment rather than a full restart. The whole cycle, including the documentation, fit inside a morning.
That result did not prove the model is ready for every commercial job. It proved that the preparation made the generation informative. Without the locked list and the revision hypothesis, the same morning would have produced attractive takes and an open question about what happens when the first real note arrives.
What to Do With the Prep Sheet After Generation
The sheet does not get thrown away. It travels with the selected take.
When the file moves to an editor or a finishing artist, the prep sheet (or a short digital version of it) answers the questions they will otherwise have to ask: what was locked, what references were used, what cleanup has already been performed, and what kinds of notes are expected to stay local.
That small transfer of context is one of the cheapest ways to protect the handoff.
The First Move This Week
If you are planning a Seedance 2.5 test, do not open the generation interface yet. Open a blank page and fill in the five sections above for one short, realistic brief. Force the locked list, the reference package, the variation budget, and one hypothesized revision note into the open.
Only then generate.
The thirty-second take will still be impressive when it works. The difference is that you will know why it worked and what happens when the next person in the chain has to change it.
Preparation is not overhead. It is the part of the workflow that decides whether the generation survives contact with the rest of the production.
If it cannot survive the handoff, it is not a workflow yet.
No letters yet — be the first to write.