Generative AI Engineering
Learn how to register, version, configure, approve, deploy, trace, roll back, and retire prompts used in production AI systems.
16 matching terms
Assets and registry
Prompt asset
A prompt and its associated metadata, variables, examples, configuration, tests, and ownership information.
Treat production prompts as governed application assets rather than anonymous copied text.
Asset includes template, schema, owner, model binding, test suite, risk level, and release notes.Assets and registry
Prompt registry or store
A controlled catalog where prompt assets can be discovered, versioned, compared, and retrieved.
Use it to prevent production prompts from being scattered across source files, documents, and personal notes.
Search the registry by use case, owner, status, model family, and environment.Assets and registry
Prompt ID
A stable identifier for a logical prompt across revisions.
Use it in code, telemetry, and approvals while versions identify exact content.
prompt_id: support.case_summaryVersion lifecycle
Prompt version
An identifiable revision of prompt content and relevant configuration.
Use immutable versions for reproducible evaluations, deployments, incidents, and rollbacks.
support.case_summary@17 with content digest sha256:...Version lifecycle
Draft and published version
A mutable working revision contrasted with an approved immutable revision available to applications.
Use drafts for iteration and publish only after required tests and approval pass.
Draft 18 remains editable; production continues to reference published version 17.Configuration
Prompt variable
A named placeholder populated with request-specific data at runtime.
Define types, required fields, escaping, length limits, and trust level for each variable.
Do not place secrets in prompt variables unless the approved design explicitly requires and protects them.
Summarize {{case_notes}} for {{audience}} using {{output_schema}}.Configuration
Prompt configuration
The model, inference parameters, tool definitions, output schema, and other settings associated with a prompt.
Version settings that can materially change behavior alongside the prompt text.
Config: model family, temperature, max output, schema version, enabled tools, timeout.Configuration
Model binding
The declared model or model family against which a prompt version is designed and tested.
Use it to prevent an untested model migration from silently changing behavior.
Prompt v17 is approved for model family A version 2026-05 only.Delivery
Environment promotion
Moving an approved prompt version through development, staging, and production environments.
Promote the same immutable version instead of manually copying text between environments.
dev -> evaluation passed -> staging approval -> productionDelivery
Prompt deployment
Making a selected prompt version and configuration available to an application or workflow.
Deploy by identifier with health checks, observability, and a documented recovery path.
Deploy support.case_summary@17 to 10% of eligible traffic, then monitor release metrics.Delivery
Prompt rollback
Returning traffic to a previously approved prompt version after a problem is found.
Prepare and rehearse rollback before production release.
If required-field accuracy falls below the threshold, restore version 16 without editing it.Governance
Prompt change approval
A documented review decision authorizing a prompt version for a defined use and environment.
Require proportionate approval for externally visible, regulated, financial, or safety-relevant behavior.
Approval records reviewer, scope, evidence, limitations, version, and expiration date.Governance
Prompt ownership
Assigned responsibility for quality, access, review, incidents, updates, and retirement of a prompt.
Assign a service owner and backup before a prompt is used in production.
Owner: Support Platform; reviewer: Compliance; next review: 2026-10-01.Governance
Prompt audit log
A tamper-resistant history of prompt changes, approvals, deployments, access, and rollbacks.
Use it to investigate incidents and demonstrate who changed what, when, and why.
Log actor, timestamp, prompt ID, before and after versions, action, ticket, and environment.Observability
Prompt trace
A request-level record linking the prompt version, inputs, model configuration, tool activity, output, and evaluation signals.
Use it for debugging and performance analysis with appropriate minimization and access controls.
Tracing can capture personal, confidential, or copyrighted material; minimize, redact, restrict, and expire it appropriately.
trace_id links prompt v17, redacted inputs, model version, latency, token use, and validator result.Version lifecycle
Prompt deprecation
Marking a prompt version or asset as scheduled for retirement and no longer recommended for new use.
Publish migration guidance, identify consumers, set a deadline, and retain rollback evidence before removal.
Deprecate v12 on August 1, migrate remaining callers to v17, and archive its evaluation record.| Record | Why it matters |
|---|---|
| Prompt ID, immutable version, and content digest | Identifies the exact asset used by an evaluation or request |
| Model, parameters, schemas, tools, and variable contract | Makes behavior reproducible and exposes compatibility requirements |
| Owner, approval, test evidence, environment, and release time | Shows authority, readiness, and operational responsibility |
| Trace policy, monitoring thresholds, and rollback target | Supports incident detection, diagnosis, and recovery |
Source control is useful, but production operation also needs prompt-specific metadata, evaluation evidence, approvals, environment promotion, request traces, ownership, deprecation, and rollback links. These can be built around code or provided by a prompt registry.
At minimum, record the settings and dependencies that can change behavior, such as model version, supported parameters, output schema, tool definitions, and variable contract. Whether they share one version number is an implementation choice, but the exact combination must be reproducible.
Do not retain unnecessary personal data, credentials, secrets, confidential source material, or unrestricted copyrighted content. Apply purpose limitation, redaction, access control, retention limits, deletion, and legal or contractual requirements.
Official guide to reusable prompts, variables, variants, testing, versions, and application integration.
Official API reference for creating and listing managed prompt versions.
Build the role, instructions, examples, constraints, reusable template, and output contract.
Generate evaluation evidence before approving and promoting a version.
Track system quality, regressions, human review, latency, and cost after deployment.
Apply privacy, access, injection defense, guardrails, moderation, and red-team controls.