epicenter

Service layer patterns: defineErrors, namespace exports, Result types. Use when: "create a service", "service layer", creating services, defining domain-specific errors.

Rust to TypeScript error handling for Tauri apps. Use when mentioning Rust errors, Tauri command errors, invoke errors, or defining Rust error types for TS consumption.

Per-change refactoring mechanics: count callers exactly, decide inline vs keep, collapse duplicate switch branches, route raw access through a single typed boundary, surgical one-change-per-commit, post-refactor straggler sweep for stale JSDoc and dead exports. Use when actually editing code to clean it up, when deciding whether a 1-caller helper earns its keep, or when planning a sequence of small refactor commits. For the smell catalog use code-audit; for session-long simplification use collapse-pass.

Use when a task seems trapped inside the current abstraction, an abstraction feels poorly designed, a fix keeps spreading across layers, or the user asks to think bigger, redesign from scratch, mentally inline, go up a level, or consider radical options. Forces a higher-level pass before coding: state the current path, invent the cleanest from-scratch option, inline suspicious layers, find asymmetric deletions, and choose the option that makes the system easiest to explain.

Query/RPC layer with TanStack Query, defineKeys, service composition, runtime DI. Use for createQuery, createMutation, queries/mutations, reactive data management.

Draft and review durable GitHub pull request titles and bodies for Epicenter. Use when creating a PR, running gh pr create, drafting or editing a PR body, writing changelog entries, linking issues, choosing merge strategy, or reviewing PR text. For local commits and branches use the git skill; for issue replies use github-issues. Never include Testing, Test Plan, or Verification sections in PR bodies unless explicitly requested.

Build a throwaway prototype to flesh out a design before committing to it. Routes between two branches: a runnable terminal app for state/business-logic questions, or several radically different UI variations toggleable from one route. Use when the user wants to prototype, sanity-check a data model or state machine, mock up a UI, explore design options, or says "prototype this", "let me play with it", "try a few designs".

Hub for the broad post-implementation review and second-read pass after an implementation: list every touched file as an ASCII tree, mentally inline helpers, audit dead paths and stale imports, name invariant owners, sanity-check API shape and naming, and delegate to focused review skills as needed. Use after finishing an implementation, before final response, or when the user says 'post-implementation review', 'review what you just did', 'second pass', or 'final sweep'.

Force one concrete sentence to find orphaned surfaces, duplicate verbs, inert abstractions. Use for "what does X actually do", "in one sentence", "too many options", or auditing whether a surface or abstraction is coherent. For a plain code-comprehension question, answer directly instead.

Monorepo scripts, package boilerplate, conventions. Use when: "how do I run", "bun run", "build this", "run tests", "typecheck", "create a new package", linting, scaffolding packages.

Method shorthand in return objects for JSDoc preservation. Use when factory functions have internal helpers that should expose docs, or hovering over returned methods shows no JSDoc.

wellcrafted/logger for library diagnostics: 5 levels, typed errors, DI sink. Use for attach primitives, background errors, file-backed logs. No console.* in library code.

Hono patterns for TypeScript API routes, middleware, request and response typing, streaming, WebSockets, and Cloudflare Workers deployment. Use when users mention Hono, honojs, Cloudflare Worker handlers, Hono middleware, or Hono route typing.

Draft a compact, cold-start prompt to paste into Claude Code, with Claude as orchestrator and Codex available through /codex:rescue for bounded evidence gathering, focused implementation, commands, verification, cleanup, and creative support. Use when the user says "hand this off", "wrap up for the next session", "resume this later", "continue in a new chat", "write a continuation prompt", "make a prompt I can copy-paste", "create a delegation brief", "prompt for an orchestrator", or invokes /handoff. Not for /goal lines (use agent-goal) or prompts that ship inside product code.

Interview the user relentlessly about a plan to reach shared understanding, resolving each branch of the decision tree. Use when stress-testing a plan or "grill me".

Greenfield clean-break review and execution for starting from an uncompromised target vision, working backward to deletion waves and owner changes, reopening settled decisions, refusing compatibility, collapsing old and new paths, moving ownership boundaries, replacing APIs, redesigning from first principles, and surfacing refusal candidates that delete disproportionate complexity. Use when the user says greenfield clean break, greenfield, clean break, no users, no compatibility burden, refuse compatibility, remove slop, collapse this, replace the API, trace upward, pressure-test the architecture, or asks whether old behavior can be deleted.

Write and edit in GOV.UK / GDS house style: plain English, active voice, front-loaded content, sentence case, and no bold or italics for emphasis. Use when the user asks for GOV.UK or GDS style, or when writing reports, research write-ups, or guidance for external readers. Not for repo docs, UI strings, commit messages, or code comments; those follow writing-voice.

GitHub issue comment guidelines for community interaction. Use when: "respond to this issue", "reply to this bug report", "close this issue", or any GitHub discussion.

Git commits and branches using Epicenter conventions. Use when staging files, writing commit messages, splitting commits, checking standalone commits, creating branches, or inspecting commit history. For pull request titles and bodies, use the pull-request skill.

Fresh-context review and adversarial review for concrete diffs, state machines, type shapes, lifecycle boundaries, and confusing abstractions. Use when the user asks for "fresh-context review", "fresh eyes", an independent reviewer, adversarial review of a diff, a state-machine audit, or whether a type or lifecycle shape earns itself. Not for generic helper delegation, executor prompts, ordinary final review, or interactive planning interviews.

First-read approachability lens. Read code as a newly onboarded TypeScript developer to surface indirection, misleading names, type-system-only tricks, and wrong-owner boundaries. Cited by post-implementation-review and collapse-pass as their readability pass; also use directly when code feels clever or hard to follow on first read.

Factory function patterns to compose clients and services. Use when wrapping resources with domain methods or refactoring mixed client/service/method options.

Error handling with wellcrafted trySync/tryAsync and toastOnError. Use for try-catch, Result types, error toasts, HTTP errors.

Drizzle ORM patterns: schema definitions, Drizzle Kit migrations, query builders, type branding, custom types, SQLite, Postgres, D1, and Turso/libSQL boundaries. Use when mentioning Drizzle, drizzle-orm, DB schemas, migrations, branded column types, or typed SQL queries.

In-code documentation, folder READMEs, code comments. Use when: "document this", "add JSDoc", "write a README", "explain this code", or writing README.md/JSDoc.