Detailed Design (per feature)

The first loop skill: it runs once per vertical slice, for the life of the project, turning a feature's planned intent into a buildable technical design — the backend/system half of the feature's low-level design (ui-design is the presentation half, and consumes this skill's contracts).

Three principles govern it:

1. The feature's what is fixed upstream; this skill designs the how. The FR/UC IDs the feature implements were assigned and verified in requirements; the plan chose and scoped the feature. This skill never invents or reinterprets requirements — that inheritance-by-ID is exactly what keeps per-feature design consistent at scale.
2. The live codebase is an input, not an obstacle. Feature N is designed months after feature 1, in a fresh session, against code that has evolved past the skeleton. Design against reality: the contracts, schema migrations, and conventions earlier features established. A design that conflicts with existing code is wrong even if it matches the documents.
3. Depth here, and only here. This is where the pipeline's depth-on-demand promise is kept: full concrete detail for this feature — contracts, columns, components, tasks — and nothing for features that haven't reached the front.

Inputs

Defaults below; user-provided paths win; source-gated citation throughout (cite FR/UC/SCR/ADR IDs only when their documents exist; never fabricate).

• Implementation plan — docs/implementation-plan.md. Which feature is next (or the user names one), its traced touchpoints: FRs implemented, UCs realized, SCR screens, building blocks/endpoints, data.
• SRS — docs/srs.md. The verbatim FR statements (they become acceptance criteria) and relevant NFRs (they constrain the design — cite them).
• Use cases — docs/use-cases.md. The UC flows the feature realizes — main/alternate/exception paths drive endpoint behavior and error design.
• Architecture — docs/architecture.md. The conceptual domain model (entity ownership, consistency boundaries), the ADRs constraining this feature, cross-cutting concepts (auth, error handling, observability conventions the design must follow).
• The live codebase — mandatory read, not optional: existing route/contract patterns, the current schema and migration history, module boundaries, shared utilities, conventions in docs/scaffold-notes.md and the agent-instructions file. Prior feature designs live in docs/features/.
• UX foundations — docs/ux-foundations.md, light touch: the SCR screens' stated purpose/states, as context for what the contracts must serve.

Outputs

Per feature, into docs/features/FEAT-NNN-<slug>/ — keyed on the feature's stable FEAT ID from the plan (immutable, so re-sequencing the plan never invalidates folder names), with the slug from its name for human readability — e.g., docs/features/FEAT-004-sign-in/:

1. technical-design.md — the feature's technical design: contracts, schema changes, component design, acceptance criteria, decisions. (Named technical-design.md, not design.md, to avoid colliding with the design system at docs/design.md; its ui-design sibling will sit beside it in the same folder.) See references/design-guide.md.
2. tasks.md — the ordered, implementable task breakdown. See references/tasks-guide.md.

Plus two write-backs:

• RTM (docs/rtm.md, if present; skip silently otherwise): append the feature's design ref (e.g., features/FEAT-004-sign-in/technical-design.md) into the Design ref cell of every FR this feature implements — append, never overwrite; Design ref column only, per the RTM ownership contract.
• Architecture amendment escalation (when triggered — see design-guide): a needed new entity or changed boundary is not designed locally; it's surfaced as a proposed architecture amendment and the feature design waits on or explicitly notes it.

Workflow

Phase 1 — Locate the feature and ingest

Identify the feature: the user names it (by FEAT ID or name — explicit naming is the only legitimate off-sequence path), or default to the next undesigned feature — computed, not stored: plan build-sequence order minus the FEAT-* folders already in docs/features/. Multiple candidates is not ambiguity — the sequence resolves it (earliest in build order wins); never present a menu. Announce the resolution ("FEAT-004 is next per the plan's build sequence") so the user can redirect before work starts. Execution status lives in the artifacts (a folder with a verified technical-design.md = designed; its tasks.md checkboxes = build progress) — the plan itself is never written with status; it stays the structural source owned by implementation-planning. Read the feature's plan entry and pull its full trace: FR IDs (fetch their verbatim statements and priorities from the SRS), UC IDs (fetch their flows), SCR IDs, blocks/endpoints, data. Read the architecture's domain model, the ADRs that touch this feature, and the cross-cutting concepts.

If the feature's plan entry is missing touchpoints (a coarse far-future feature that just reached the front), tighten it first: derive the missing FR/UC/SCR trace from the documents and confirm with the user — this is the plan's just-in-time elaboration happening on schedule.

Phase 2 — Read the code

Survey the live codebase before designing anything: existing endpoint and contract conventions (naming, envelope shapes, error format), the current schema and its migration history, the module the feature lands in and its boundaries, shared utilities the design should reuse, and prior feature designs in docs/features/ (especially any this feature depends on). Record what the design must conform to. Reality wins over documents where they diverge — and note the divergence.

Phase 3 — Design

Read references/design-guide.md. Produce, concretely:

• API contracts — endpoints with methods, paths, request/response shapes, validation rules, error codes; following the codebase's established conventions. UC exception/alternate flows become error behaviors.
• Physical schema changes — tables/columns/indexes/constraints as migrations, within the entities the architecture's conceptual model already owns. A new entity or ownership/boundary change escalates to an architecture amendment — never invented locally.
• Component design — the modules/services/functions inside the feature's building block: responsibilities, interactions, where the stubbed skeleton code gets replaced.
• Acceptance criteria — testable, derived from the FR statements and UC flows, each citing its FR/UC ID. These are what acceptance verification will later check.
• Design decisions — any real fork gets a short decision note with its driver (mini-ADR discipline; feature-local decisions stay here, architectural ones escalate).

Phase 4 — Break down tasks

Read references/tasks-guide.md. Decompose the design into tasks.md: an ordered list of implementable tasks, each small enough to complete and verify in one focused session, each pointing at the design sections and acceptance criteria it serves, ordered so the feature builds incrementally (schema → domain → contract → wiring → tests green), with UI tasks referenced but owned by ui-design's output.

Phase 5 — Verify

Read references/verification.md and self-check: every feature FR is covered by acceptance criteria and by tasks; every cited ID resolves; the schema stays inside the conceptual model (or an escalation is filed); contracts don't collide with existing code; tasks are complete, ordered, and traceable. Fix failures; flag the unfixable.

Phase 6 — Deliver

Write both files, do the RTM append, and summarize: the design in a few sentences, the task count and order, any escalations filed, a live-computed progress line ("FEAT-004 designed — 5 of 14 features now designed", counted from the plan vs. docs/features/; never stored anywhere), and the handoff — ui-design now designs the feature's SCR screens against these contracts; implementation follows tasks.md. Offer to proceed to either.

Scope boundaries

Does not: design screens (ui-design's half — it consumes these contracts); implement anything (tasks.md is the handoff to implementation); change requirements or the plan (amendments belong to their owning skills); invent entities or alter consistency boundaries (architecture amendment); write to any RTM column except appending Design ref.

What good looks like

• Every acceptance criterion cites the FR/UC it verifies; every task points at design sections; nothing in the design lacks a requirement behind it.
• The schema diff is migrations against the actual current schema, not the skeleton's — and it stays inside the conceptual model's entities.
• Contracts read like the codebase already reads — conventions inherited, not reinvented.
• tasks.md could be handed to a fresh session (or a loop agent) and executed top-to-bottom without re-deriving context.
• Escalations were filed, not worked around.