When several renders look similar, it is easy to forget which prompt produced the useful one. A version log protects the learning from each attempt. It is especially valuable when a team revisits a project after the service or model has changed.
Record the complete experiment
Store the source filename, prompt text, model or workflow label shown in the interface, exposed settings, date, and output filename. Include the cost shown for the attempt when available. A prompt alone is not a reproducible recipe: source material and settings can matter, and a service may change behavior over time. The log preserves evidence, not guaranteed identical regeneration.
Change one variable when diagnosing a failure
If you change the prompt, source clip, and strength together, a better result does not tell you which change helped. For a targeted experiment, keep the source and other controls fixed and change one relevant factor. For open-ended creative exploration, multiple changes can be reasonable, but label that attempt exploratory rather than a controlled comparison.
Keep rejected outputs long enough to learn
Add a short rejection reason and timestamp instead of deleting every unsuccessful result immediately. Group failures by type: wrong action, identity drift, framing, or technical export. These patterns can guide the next experiment or a decision to stop. Follow your own retention and privacy rules for footage; the log should not become an uncontrolled archive of sensitive material.
Decision worksheet
| Item | Record or check | Decision rule |
|---|---|---|
| Input | Source clip and version | clip-B-v01 |
| Instruction | Exact prompt and changed variable | background description only |
| Output | Filename and visible workflow label | render-007 |
| Decision | Accepted seconds and rejection note | 4 seconds; drift after 00:04 |
Worked example — illustrative, not measured
Illustrative test: attempts 06 and 07 use the same clip and settings, but 07 removes a conflicting camera instruction. If 07 preserves the action better, record that observation narrowly. Do not conclude that removing camera instructions always improves every model; repeat or test another clip before generalizing.
Put this into practice
- Create the log before the first paid attempt.
- Give each attempt a unique identifier.
- Save the exact prompt rather than a paraphrase.
- Review the pattern of failures before buying more capacity.
A mistake to avoid
Do not overwrite the only accepted file with a new render. Keep accepted outputs immutable and create a new version for every experiment.
Basis and limitations
Methodology: this is an original editorial planning framework, not a vendor feature claim or a hands-on product test. Adapt the suggested checks to your project.