Glossary
The vocabulary of AI production memory.
Clear, quotable definitions of the terms behind Behind The Workflow and the AI creative work it documents, from production memory and provenance to prompts, references, verdicts and the EU AI Act.
- Production memory
- Production memory is the full record behind a piece of AI creative work, kept in one searchable place. It holds every asset together with its prompt, tool, model, references, versions, decision, author and time, so the work stays findable, reproducible and ready to hand off long after it was made.
- AI provenance
- AI provenance is the documented account of how an AI asset was made: the prompt, model, references, decision, author and time that were declared and attached to it. In BTW this account is structured and reproducible, and it describes what was recorded rather than serving as cryptographic proof. Cryptographic content credentials are a separate mechanism, described under C2PA below.
- Prompt documentation
- Prompt documentation is the exact text that generated an asset, kept verbatim and searchable alongside the result. Keeping the prompt on the asset is what lets anyone read back the instruction that produced a shot and run it again.
- Model and seed
- The model is the specific generator and model version that produced an asset, such as a named image or video model. The seed is the starting value that, together with the prompt and settings, lets the same result be generated again. Recording both is what turns a single output into a repeatable one.
- Asset lineage / version history
- Asset lineage is the connected chain of versions behind a single shot, held in the order they were made. Instead of scattering variations across folders, the whole line of attempts stays attached to the shot, so the path from first draft to final stays visible.
- Reference (rights status: clear / unclear / risky)
- A reference is an image or input attached to a shot to steer a generation. Each reference carries a rights status that marks whether its use is clear, unclear or risky, so the team can see at a glance which inputs are safe to ship and which need a second look.
- Verdict (hold / kill / final)
- A verdict is the recorded decision on a version: hold it for later, kill it, or mark it final. The verdict sits on the shot next to the version itself, so weeks later nobody has to guess which version shipped and why.
- Client handoff
- A client handoff is the moment a finished asset leaves the team with its full record attached: the prompt, model, references, decision, author and time. The person on the other end receives the picture and the account of how it was made together, in one share or export.
- EU AI Act Article 50
- Article 50 of the EU AI Act sets transparency and documentation duties for AI-generated content, which begin to apply on 2 August 2026. It expects makers to be able to show that content was AI generated and to document how it was made. BTW produces that documentation by default, as a byproduct of the normal workflow.
- C2PA / content credentials
- C2PA, also called content credentials, is an open technical standard that attaches signed, tamper-evident metadata to a file using cryptographic signatures. It is a separate mechanism from a declared production record: where a declared record documents what was stated and attached, content credentials add a signature layer that can be verified. The two complement each other rather than replace one another.
- Reproducibility
- Reproducibility is the ability to recreate an asset from its recorded recipe. When the prompt, model, seed, references and settings all stay on the asset, a result can be run again instead of reconstructed from memory.
- Generative asset
- A generative asset is a piece of media produced by an AI model, such as an image, a video or a clip. In production memory it is never just the file: it is the file together with the recipe and decisions that produced it.
- Prompt library
- A prompt library is the searchable collection of prompts a team has actually run, kept with the assets they produced. It turns scattered, one-time instructions into a reusable resource, so a phrasing that worked can be found and run again.
- Digital asset management (DAM) in the AI context
- Digital asset management, or DAM, is the practice of storing and organising finished media files. In the AI context this is not enough on its own, because the value sits in how each asset was made. Production memory extends DAM by keeping the prompt, model, references, versions and decisions attached to every asset, not just the final file.
Get started
Stop rebuilding work your team already did.
Set up your first project in about two minutes. From day one, every asset, prompt, reference and decision stays findable.