Four views of one system. The goal: separation of responsibilities, not more process layers. Each building block has one job; whatever is shared across roles lives in one place — never in copies.
An agent skill contains only the behavior of its role or domain. How to talk to a human, hand work between agents, access tools and record evidence — those are shared contracts, defined once and shared by every role.
says which stages exist and when work moves forward.
say who (which capability) performs the specialist work.
says how each role talks to a person — one shared contract.
says how agents hand work to each other — a concise, structured format.
says how deterministic packets are encoded — only where needed.
say when an agent may decide, must ask, must stop.
define how authoritative systems are read and changed.
supplies the truth about the domain: specifications, ADRs, security requirements.
prove what really happened — not what was declared.
The process and skills depend on shared contracts — never the other way around. If behavior is shared by multiple skills, it moves into a shared contract instead of being copied. Copies always drift apart; one contract does not.
Target rule: if behavior is shared by multiple skills — move it into a shared contract instead of copying it. No skill may define a competing communication contract.
Every unit of work goes through the same loop: understand the goal, pick the owner, specialist work on authoritative systems, evidence collection, gate. The human sees goals, decisions, risks and outcomes — internal execution packets stay inside until they ask for them.
All delivery rides one line. Required controls — security, architecture — join it as supporting evidence, they do not create detours. Every transition between stages has a named exit gate and exactly one role that fires it. Jira holds the truth about the stage, GitHub the truth about the code, and the human conversation is about goals and decisions — not every intermediate station.
This is a map of responsibilities, not a second process engine. Detailed stage rules live in the canonical process documentation, role behavior in skills, shared communication in shared contracts. The map shows where each thing lives — and why nowhere else.
Every change goes through the same lifecycle — but the required evidence bundle scales with the change's reach. The classification is verified by a gate based on what the change actually touched, not on the author's declaration.
| Reach | What it covers | Required evidence bundle |
|---|---|---|
| LOW | Documentation, descriptive configuration — changes that do not alter system behavior. | Lint, page/document build, one review. No full product test run. |
| STANDARD | Product code — logic, APIs, interface, tests. | Full bundle: tests, lint, formal review, QA verdict pinned to the exact revision. |
| HIGH | Authentication, cryptography, data migrations, infrastructure. | The standard bundle extended with a mandatory security opinion; all verdicts bound to the same revision. |
What this is not: these are not three process paths. There is one path — the same sequence of stages leads to "Done". Only the depth of required evidence scales, and the reach classification is verified by a gate, not by the author of the change.