The guardrails · Engineering
Engineering Guardrails
The rules AI builds inside, and the customer prompt
AI writes most of the code. Engineering owns the rules it writes inside. Each guardrail is written once into the context files, so AI follows it every session, and enforced by the Pipeline, so nothing ships without it.
Context files · review
Architecture Boundaries
AI builds inside the lines. Approved services, patterns and layers live in the context files – what AI may change, what it must never touch, and how parts talk to each other. Engineering reviews any change that crosses a boundary.
Pipeline · review
Data Governance
Customer data stays where it belongs. Data is classified as public, internal, personal or sensitive. Personal data never enters prompts, logs or test fixtures. Retention and residency rules live in the context files.
Pipeline · final review
Model Safety
AI features behave, every time. Any AI feature customers touch has evaluations for accuracy, tone and harmful output, run on every change – plus protection against prompt injection, and a human fallback.
Pipeline gate
Performance Budgets
Fast is a feature. Budgets for page load, response time and size, set per product area. A change that breaks a budget doesn't merge.
Before it ships
Observability
If you can't see it, it doesn't ship. Every Intent ships with logs, traces, error alerts and its measure instrumented on the Dashboard – so Thursday's KPI Learnings, and any rollback, have real data.
Pipeline gate
Security Gates
Checked on every merge. Static analysis, secret scanning, dependency vulnerability scans, and least-privilege access for AI agents. A failed gate blocks the merge.
Pipeline · review
Dependency Management
Nothing new without a reason. New libraries need Engineering approval. Versions are pinned, licenses are checked, and every AI-suggested package is verified to exist before it's installed.
Who owns them. Engineering writes and maintains the guardrails. The Pipeline enforces them automatically. Final Review confirms them for anything that touches customer data, payments or logins.
Customer Prompt Guardrails
The rules for the "Was this better?" prompt and the First Circle.
Ask Once
Once per customer per Intent. Never on every visit.
Never in the Way
After the task is done, never in the middle of it. Easy to dismiss.
Opt-In Only
The First Circle is always an invitation. Respect privacy and consent.