Teams nod along to production memory, then ask the practical question: what exactly do we store? The AI production record is the answer. It is a concrete, per-version document — structured enough for systems, readable enough for humans — that carries everything a future contributor needs. If production memory is the library, the production record is the book.
Anatomy of a production record
- Identity: which canonical asset and which version state this record describes.
- Inputs: brief link, source references, style guides, licensed material.
- Generation context: full prompt, negative prompts, seed, model and tool versions, parameters.
- Cause: what changed since the previous version and why.
- Status: draft, in review, approved, or superseded — with author and timestamp.
- Decisions and approvals: reviewer notes, sign-off rationale, constraints.
- Lineage: parent version, sibling variants, and downstream derivatives.
Example record
Real implementations store this as structured data. The AI asset provenance guide shows one way to represent it; the discipline matters more than the format. A complete record in a simple schema beats a rich schema nobody fills.
Machine vs. human fields
Who records what
| Field | Captured by | When |
|---|---|---|
| Model, parameters, seed | System, automatic | At render |
| Prompt and references | System + artist | At creation |
| Revision cause | Artist | At version creation |
| Approval and rationale | Reviewer | At decision |
| Lineage links | System | At branch and delivery |
Implementation checklist
Production-record checklist
- Fix the schema: the seven sections above, no more to start.
- Automate machine fields so completeness is the default.
- Require cause on every new version and rationale on every approval.
- Link lineage at branch time, not in retrospectives.
- Audit monthly: sample ten records, score completeness.
What to measure in a paid pilot
Track record completeness, reproduction success using records alone, and handoff time for record-backed versus record-less assets. See the production pilot page to scope a pilot.
Start with the workflow assessment to find where your team loses context today, read the Production Memory guide, and browse the use-case library for related workflows.
Key takeaways
- One record per asset version: identity, inputs, context, cause, status, lineage.
- The record is the contract between everyone who touches the asset.
- Machine-captured fields plus human-entered decisions — neither alone suffices.
- A concrete record schema beats abstract metadata advice every time.