Cross-platformWorkflow & organisation
How to Track Prompts Across AI Image and Video Tools
Keep one record per output with eight fields: platform, model or version, prompt, settings, references used, date, status and where the file lives. A dated folder per project plus one spreadsheet is enough. Local tools such as ComfyUI and A1111 embed the prompt in the image file; most web tools do not, so you must copy it at the moment you generate.
Anyone who uses more than one AI image or video tool eventually hits the same problem: a client asks for 'that version from last Tuesday, but with the blue background' and nobody can remember which platform made it, let alone the prompt. The fix is not a complicated tool. It is a habit of writing one record per output at the moment you make it, in a format you can search later. This guide gives you the fields, a folder-and-sheet system that works across every platform, and a note on which tools save metadata for you and which do not.
What to record for every output
Each generated image or clip you might ever want again deserves one row. Eight fields cover almost every case.
- Platform. Midjourney, Kling, Flux on a hosted service, ComfyUI locally, and so on.
- Model or version. The exact model name and version if the interface shows it, such as a Gen-4 variant, a Kling Professional mode, or a checkpoint filename. Models change monthly and the same prompt behaves differently on each.
- Prompt. The full text, including negative prompts and any parameters embedded in the prompt line.
- Settings. Aspect ratio, duration, resolution, seed, stylisation values, guidance, steps, camera preset, and anything else you changed from default.
- References. Filenames or links of any image, style reference, Element, Ingredient or audio you supplied.
- Date. When it was generated, so you can match it to the model that existed that day.
- Status. Draft, shortlisted, approved, delivered, rejected. This is the field clients care about.
- Location. Where the file is saved and its filename.
If that sounds like a lot, note that five of the eight can be copied and pasted in seconds. The discipline is doing it before you click generate again.
Which tools save the prompt for you
| Tool type | Examples | Prompt stored in the file? | What you must do |
|---|---|---|---|
| Local open-source UIs | ComfyUI, A1111 (Automatic1111), Forge | Yes, usually as PNG metadata (the workflow or the generation parameters) | Keep the original PNG; converting to JPEG or uploading to most services strips it |
| Hosted image tools | Midjourney, Flux hosted, Ideogram, Leonardo, ChatGPT images, Gemini | Mostly no; the prompt lives in the web interface or chat history | Copy the prompt and settings into your record as you generate |
| Hosted video tools | Kling, Runway, Luma, Pika, Google Flow, Higgsfield | No; the exported video carries no prompt | Record prompt, model, duration and references per clip; name the file to match the row |
| Hosted model playgrounds | Replicate, fal | Sometimes a run log in the web interface, not in the file | Copy the input JSON into your record |
Web histories are convenient until they are not: they are per account, they can be pruned, and they are useless when the account belongs to a freelancer who has moved on. Treat them as a backup, not the record.
A folder-and-sheet system
You need two things: a predictable folder structure and one spreadsheet per project. Keep both in the same shared drive the rest of the project uses.
Folders
ProjectName/at the top.01-exploration/,02-shortlist/,03-approved/,04-delivered/for status. Move files forward; never copy, so there is one version of truth.refs/for every reference image, audio file and style sample you fed to any tool.- Filenames that carry the key facts:
2026-03-14_kling-pro_bottle-dolly_v3.mp4reads as date, platform and model, subject, version.
The sheet
One row per output, one column per field above, plus a free-text notes column for what went wrong or right. Add a filter on status and a filter on platform, and you can answer 'show me every approved Midjourney image' in a click. For video, add a column for the source still, because the still is usually what you will reuse. If your team already lives in a project tool with custom fields, the same eight fields work there too.
Habits that make it stick
- Record before you iterate. The moment you see a result you might keep, paste the prompt. Iterating first means losing the version you liked.
- Copy the whole prompt, not a summary. Paraphrases cannot be re-run.
- Save references with the output. A Runway References shot or a Kling Elements clip cannot be reproduced without the reference images.
- Note the model when a platform updates. If a tool announces a new version, add a row in your notes so you know why older prompts changed behaviour.
- Keep PNG originals from local tools. The embedded workflow is the most complete record you will ever have, and it is gone once you export to JPEG.
For client work, this record is also your evidence. Our guide on keeping a record of AI work for clients covers what to hand over and why.
Try it
Set up your sheet, then generate this prompt on two different platforms and fill a row for each. The point is the habit, not the picture.
A single green apple on a weathered wooden table, soft side light from a window, shallow depth of field, muted colours, photographic, no text
Change one setting on the second run, such as aspect ratio or a style value, and record it. Then close both tools, open your sheet and try to reproduce the first result from the row alone. Any field you could not fill is the one your system is missing.