Pre-launch · Training and the free Foundations exam are open now. Levels 1–5 open soon. Get notified on any level page.
20 min read
A release is a decision someone is accountable for. At enterprise scale, one shared prompt or schema may serve dozens of workflows, so a change can reach far beyond its author's view. This module covers the full lifecycle: designing contracts and components that can evolve, releasing changes in stages with evidence, keeping decisions reconstructable, and retiring old versions so nothing depends on them by accident.
Shared components need typed schemas with explicit unknowns, so a missing value is never mistaken for zero. Classify each change by its behavioral effect, not its size: a one-word edit that changes who gets escalated is a breaking change. Adding a field breaks any consumer that rejects unknown fields. Version contracts, test old and new consumers, and schedule migrations. Domain teams can adapt a shared core, but shared controls stay fixed and each domain keeps its own test suite.
A model proposes a purchase or refund. Trusted logic checks identity, limits, and policy, and a person signs off where required. Execution uses a durable request ID so a retry after a timeout can't pay twice, and every step is logged.
Multi-stage workflows give each stage one role and a validated handoff. Claims carry source, version, location, and retrieval date. Expired or unsupported claims cause abstention or escalation. Rationale is a checkable list of criteria, evidence, and assumptions, not a claimed transcript of model thinking. Store enough lineage that, a year later, you can reconstruct what the system saw and why it recommended what it did.
Map the blast radius: every workflow that uses the component. Gate model migrations on evidence from real workloads and high-risk slices, not generic benchmarks. Roll out in stages with monitoring and a named owner who can roll back. When retiring a version, inventory its consumers, confirm each has moved, and verify through logs that no calls remain.
Olivia Brandt owns the shared requisition component at a manufacturer with five divisions. It turns requests into purchase proposals. She needs to add a hazmat_class field and move to a new model. Two divisions' systems reject unknown fields.
Weak approach. Ship v3 to all divisions on Monday with the new model, since benchmark scores are higher.
Better approach.
Before:
Release v3 + new model to all divisions. Benchmarks improved.
After:
Release plan: requisition-core v3.0 (owner and rollback authority:
O. Brandt)
Change class: breaking (new field; escalation wording changed).
Blast radius: 5 divisions, 14 consumer systems; 2 reject unknown
fields.
Contract: hazmat_class enum or "unknown"; amount_usd number;
purchase still requires rules-engine check and buyer sign-off.
Evidence gate: new model on 1,200 real requisitions per division
plus 150 oversampled hazmat cases; hazmat recall must be at least
99%. Overall estimates reweighted to real hazmat share (2%).
Rollout: Divisions A, B (week 1) -> C (week 2) -> D, E after
consumer upgrades. Rollback trigger: any hazmat miss or schema
rejection.
Fallback: prior model may draft only; cannot flag hazmat as
"none"; users see "limited mode."
Lineage: request ID, model and contract version, sources stored
3 years.
Retirement: v2 sunset after logs show zero v2 calls for 30 days;
consumer owners confirm in writing.
Why this is stronger. The change was classified by its effect, and incompatible consumers were found before release. Migration was gated on real and high-risk evidence. Rollout was staged with a named owner and triggers. Fallback was constrained and disclosed. Retirement was verified, not assumed.
Setup. Pick a shared prompt, template, or schema that more than one team uses.
Task. List every consumer you can find. Classify one recent or planned change as compatible or breaking, and explain why. Write a three-stage rollout with one rollback trigger and a retirement condition for the old version.
What a good result looks like. Your consumer list includes at least one you didn't expect. Your classification is based on behavior, not edit size. Your retirement condition can be verified from logs.