magnus919_agent-skills/forward-deployed-engineering/references/route-selection.md

Specialist Route Selection

The stage lists in manifest.yaml are candidate routes, not an instruction to load every listed skill. Load one primary specialist when its entry condition is true. Add a secondary specialist only for a named blocker, risk, or handoff. If one specialist fully owns the request, leave the FDE lifecycle and route the bounded task directly.

Stage	Candidate	Load only when	Do not load when
Discover	product-discovery	The user, workflow, problem, outcome, or stakeholder interpretation is unvalidated	The problem and acceptance boundary are already explicit
Discover	data-engineering	Discovery depends on data sources, contracts, quality, lineage, or pipeline feasibility	Data is incidental to the workflow question
Discover	remote-systems-administration	Relevant evidence or action sits behind remote access, privilege, egress, bastion, or recovery constraints	No remote or constrained system is involved
Frame	product-design-and-ux	The frame depends on user journeys, interaction behavior, service experience, or usability constraints	The work has no user-facing or workflow-experience question
Frame	privacy-engineering	The frame includes personal data, consent, retention, disclosure, or privacy rights	No privacy-relevant data or processing is in scope
Frame	secure-software-engineering	The frame requires threat, trust-boundary, authentication, authorization, or secure-design decisions	Security is not a material part of the proposed intervention
Hypothesize	product-experimentation	The intervention needs a causal or comparative product/workflow experiment	The hypothesis is verified through engineering acceptance rather than an experiment
Hypothesize	agent-evals-and-observability	An agent, LLM, model, or nondeterministic workflow needs a decision/risk contract and evaluation plan	The capability is deterministic and ordinary test methods are sufficient
Build	implementation-planning	An authorized decision-maker has approved the requirement or specification and a dependency-aware delivery plan is needed	The requirement is not approved; stop and return upstream
Build	backend-engineering	The approved slice requires backend service or API implementation	No backend component is in scope
Build	frontend-engineering	The approved slice requires a web frontend or client implementation	No frontend component is in scope
Build	data-engineering	The approved slice requires data pipelines, contracts, storage, or transformation	Data work is not part of the implementation
Evaluate	agent-evals-and-observability	Agent, LLM, model, or nondeterministic behavior needs representative, adversarial, trajectory, or production-feedback evaluation	Ordinary deterministic verification is sufficient
Evaluate	qa-methodology	A risk-scaled test strategy or coverage model is needed	Acceptance criteria can be verified directly without a broader QA design
Evaluate	verification-methodology	Explicit acceptance criteria need direct pass/fail evidence	Criteria are not yet defined; return to framing or hypothesis work
Deploy	production-readiness	A production or high-impact release needs a risk-scaled readiness verdict	The change is not crossing a production-like boundary
Deploy	release-engineering	Release mechanics, versioning, promotion, rollback, or delivery automation are needed	No release artifact or promotion path exists
Deploy	platform-engineering	Deployment is blocked on an internal platform capability or platform-owned interface	The engagement itself is ongoing platform ownership; route directly instead
Deploy	remote-systems-administration	Deployment occurs through a constrained remote access or recovery path	The environment is local and unconstrained
Adopt	product-adoption	Activation, onboarding, feature discovery, behavior change, or sustained use is below the decision rule	Intended workflow adoption is already evidenced
Adopt	product-design-and-ux	Adoption evidence points to usability or workflow-experience friction	The barrier is access, support, incentive, or ownership rather than UX
Measure	product-analytics-and-measurement	Outcome metrics, instrumentation, events, funnels, cohorts, or governance are needed	Existing measurement evidence already answers the decision rule
Measure	agent-evals-and-observability	Model or agent quality needs production feedback and evaluation drift evidence	No agentic or nondeterministic component exists
Measure	site-reliability-engineering	Reliability is an agreed outcome or a blocker to workflow impact	The request is ongoing SRE ownership; route directly instead
Generalize	product-lifecycle-learning	Expected and observed outcomes must be compared to choose continue, improve, pivot, pause, or retire	There is no post-launch evidence to compare
Generalize	product-methodology	A bounded classification, prioritization, or decision log is needed for the local pattern	The question is portfolio investment or full product lifecycle governance

Approval boundaries

• implementation-planning is never a pre-approval route. A hypothesis, prototype, or promising field pattern is not an approved requirement.
• product-operations-and-governance is not a Frame-stage route for a single engagement decision. Load it only outside this bundle when the request is to design a recurring governance system.
• A Generalize-stage classification may hand an approved requirement to implementation-planning as the next owner's work. The planning skill is not part of the Generalize decision itself.