Practical guide

Working with CRDs

The operating guidance that keeps a CRD authoritative without turning it into a mini-PRD.

Decision precedence

When applicable guidance conflicts, resolve it in this order:

1. Rule / invariant
2. Explicit implementation requirement
3. Explicit owner or user choice
4. Recommended default
5. Agent judgment

Lower-precedence guidance cannot override higher-precedence guidance. Same-level conflicts or unclear scope are unresolved decisions, not invitations for an agent to improvise.

Shared elements and approved reuse

Reusable code and tools are not automatically reusable capability meaning. A shared element may retain assumptions from the capability it was created for.

shared element: customer-identity-validation
created for: customer-onboarding
approved reuse: [request-moving-quote, submit-insurance-claim]

Use this optional declaration for components, prompts, schemas, workflows, tools, or tests when origin context and explicit reuse approval matter.

Capability, skill, and tool

ConceptMeaningTypical relationship
Capability MLEComplete contextual ability that produces an outcome.Defined by one CRD.
SkillAgentic realization package: instructions, context, tools, and workflow.May realize one capability or bundle several.
ToolExecutable primitive: API operation, MCP tool, function, command, or workflow step.May support many skills and capabilities.

A skill is not automatically a capability, and a tool is not automatically a skill. A capability can be the MLE of a skill only when the skill has one coherent outcome.

Related MLEs by Dimension

A CRD may optionally trace itself to lower-level, discipline-specific MLEs — Business/Domain, UX/Experience, Communication, Interaction/Behaviour, Frontend/Interface, Backend/Execution, Data/Information, API/Interoperability, Agentic, Verification, Operations — using relationship types defines, implements, supports, constrains, verifies, exposes, or reused_by (full model: specification §7.4). This is traceability, not coverage: never fill a dimension with no genuine content just to appear complete.

Communication is deliberately first-class here rather than filed under UX, because a message's meaning is usually shaped by business rules, rendered by frontend or delivery infrastructure, and localized independently — no single discipline owns it. A Communication MLE separates canonical meaning from its channel realization and its language realization, so the same intended meaning stays consistent across a toast, an email, a push notification, and a translation of any of those.

One capability-owned Communication MLE and one product-wide Source Context Reference are complementary, not competing: the message itself belongs to the capability that produces it, while a cross-cutting terminology or tone convention many messages should follow belongs in SCR if the product already has one. See the worked trial in Reconcile payments.

Why CRD is not merely a feature or existing artifact

ArtifactUseful forWhat it does not establish alone
PRDProduct/application rationale and scope.A reusable, independently bounded capability.
User storyA human actor’s desired outcome.Rules, state transitions, machine contracts, and realization independence.
UI screen or Storybook componentAn interface representation.Complete capability purpose, policies, effects, and non-UI use.
API/OpenAPI/GraphQL moduleTechnical interface and data contract.Whether operations together form one meaningful capability.
Skill or MCP toolAgentic realization or executable interface.Purpose, boundaries, and approved capability context.

“Feature” is a useful informal word, but its boundary is often negotiated case by case. A CRD makes the boundary explicit through purpose, outcome, included/excluded concerns, contracts, rules, defaults, and unknowns.

Audience projections

One canonical CRD may be projected for different readers without duplicating or diverging its requirements.

ProjectionQuestions it answers
Business/domainWhy does it exist? What outcome, rules, and rationale matter?
UXWho acts? What choices, states, and outcomes must be understandable?
FrontendWhat representations and interaction constraints are needed?
BackendWhat transitions, invariants, effects, and dependencies must hold?
API/MCP/toolsWhat commands, contracts, inputs, outputs, and authority bounds exist?
AgentWhat authority, grounding, precedence, approval, and escalation rules apply?
OperationsWhich realization is active and how is it observed, recovered, and retired?

A projection is a view of the CRD, not a new source of truth.