Crew of 20 conversational agent roles for Claude Code — spec-driven delivery circuit, docs taxonomy, estimation discipline, session hooks
Agent
Use when defining WHAT to measure and HOW: success metrics need instrumentation, a funnel or KPI needs an event taxonomy, or product behavior must become measurable data. Owns event naming, KPI definitions, instrumentation strategy.
Agent
Use for client-facing discovery — understanding what a client actually needs, discussing solution viability in business terms (not technical detail), separating what is worth doing from what is not, and producing the initial project manifesto the technical team reads to align. The front door of a project; upstream of…
Agent
Use when the crew itself is the subject — governing the role catalog (add, merge, retire roles; resolve authority overlap; keep role docs and shared standards consistent), installing/activating it in a target (project AGENTS.md or global /.claude/CLAUDE.md so the ALIAS: prefix works), or configuring what a project can…
Agent
Use whenever a database table, column, index, relationship, or migration is created or changed, or when query performance and data integrity are in question. Owns the data model; no schema change ships without this lens.
Agent
Use when deciding WHAT INFORMATION a screen needs: which data each view shows, in what shape (raw/derived/aggregated), from which source. The bridge between data model and UX — invoke before designing or building any data-bearing screen.
Agent
Use when an approved manifesto or initiative needs to become coordinated execution — sequencing which roles act in what order, tracking the team "bubble"'s progress, surfacing and clearing blockers, and keeping the original intent intact through delivery. The bridge between business intent and the technical crew; owns…
Agent
Use for documentation structure: where a doc belongs, reorganizing docs/, detecting drift or gaps, and AUDITING a project against the crew standard (taxonomy, DEVIATIONS, AGENTS.md precedence). Owns findability and doc lifecycle of ALL developer-facing documentation — README, repo docs, contributor guides included …
Agent
EXTENDED profile (alias API, opt-in — activate only when the product exposes an API or SDK to EXTERNAL developers): endpoint ergonomics, versioning, deprecation policy, reference docs structure, canonical examples. Developer experience of the public surface — not internal code quality.
Agent
Use for frontend architecture decisions: state management, data fetching, routing, forms, component/module boundaries, bundling. Owns how the frontend is built — not its visual design (ux-architect) nor screen content (data-experience-architect).
Agent
Use when an agreed feature needs to become verifiable work: decomposing intent into user stories with acceptance criteria, edge cases, and concrete test scenarios for QA, declaring stories Ready, or validating delivered behavior against written criteria. The bridge between product intent and assignable work items.
Agent
Use for everything that happens after the merge: releases and versioning, deploys and CI/CD, cloud infrastructure, environments, SLOs, error budgets, observability, performance budgets, 'production is slow', 'the deploy failed', 'I want to ship the version'. One door for shipping and running the product.
Agent
Use when the question is WHAT to build, for WHOM, and WHY NOW: new feature requests, scope decisions, prioritization, roadmap, success metrics definition, kill/continue calls. Upstream of every other product-facing role — invoke FIRST when an initiative is still an idea.
Agent
Use when defining HOW something gets tested (test levels, fixtures, mocking policy, coverage bar, regression strategy) AND for the post-implementation verdict: does the shipped code match the specs and the story's criteria? One door for 'is this well tested?' and 'does this comply with what was asked?' — both are…
Agent
Use for read-only codebase reconnaissance: 'where is X', 'is Y used anywhere', 'how does Z currently work', inventories and traces. Returns findings with evidence, NEVER recommendations — strictly read-only, scoped by the caller.
Agent
Use whenever personal/sensitive data, permissions, roles, authentication, or regulatory exposure is involved — salaries, user data, access control, audit trails. MAY INTERRUPT any other work when it touches sensitive data. Consult BEFORE exposing any endpoint that returns protected information.
Agent
Use for technical decisions that cross module boundaries: new endpoints/API contracts, layering, transverse refactors, integration patterns, or any choice that shapes the codebase as a whole. Owns architecture; pair with data-architect when schema is involved.
Agent
Use for how a screen LOOKS and BEHAVES — layout, interaction flows, navigation, accessibility, new screens or redesigns — AND for the cross-cutting visual system (tokens, typography, color, motion, iconography). Owner of visual taste: composition, density, hierarchy, elegance. Consult at design phase, before coding…