Prompt Quality Gate

Version: 1.0.0
Last updated: 2026-07-16

Purpose

Provide a verifiable release gate for production prompt changes.

Why

Prompt review by inspection misses regressions, unsafe authority assumptions, and distribution-specific failures. A prompt is complete only when its contract and evaluations pass.

How

Complete every required item. A failed security, authorization, privacy, schema, or regression item blocks release; do not use an aggregate passing score.

Contract

• Role, context, instructions, output format, and constraints are separated
• Task scope and abstention behavior are explicit
• Output has a machine-checkable schema where consumed by code
• No instruction requires exposed chain-of-thought
• Examples are representative, consistent, and free of sensitive data

Security and Data

• Untrusted input is identified as data, not authority
• Identity, authorization, and side-effect policy are enforced outside the model
• Retrieved and tool-provided content is treated as untrusted
• Only required, authorized data is submitted
• Logs and retention comply with data classification

Evaluation

• Representative, edge, adversarial, and abstention cases pass
• Baseline comparison shows no material regression
• Quality, latency, and effective cost thresholds pass
• Model and prompt versions are pinned in results
• Rollback criteria and owner are recorded

When

Run before initial release and after any prompt, model, tool, retrieval, schema, or policy change.

Tradeoffs

Benefit	Cost
Repeatable release evidence	Evaluation maintenance
Security controls stay explicit	More integration work
Faster rollback	Versioning overhead

Anti-Patterns

• Percentage pass scores that permit a critical control to fail.
• Manual spot checks as the only evaluation.
• Treating injection phrase detection as authorization.
• Approving prompts independently of their harness.

Enterprise Considerations

Store gate results with change records, control mappings, approver identity, model/version, dataset lineage, and retention policy. Require independent review for regulated or high-impact use cases.

Checklist

• All blocking sections above pass
• Evidence is attached to the release record
• Named owner accepts residual risk
• Production monitoring covers gate metrics
• Rollback is tested

Changelog

• 1.0.0 (2026-07-16): Initial risk-based prompt release gate.