Vibebuilder

Explicit GPT-5.6 native-first router for auditing, designing, installing, or validating Codex skill combinations, global skill setup, routing fixtures, and harness behavior. Use when the user explicitly invokes $codex-skill-router or asks to change or inspect Codex skills, routing, AGENTS guidance, or global harness configuration. Do not use for ordinary coding, review, debugging, research, or UI work merely because the task is complex.

Use when work introduces or changes shared nouns, states, IDs, ownership, authority, or customer-facing vocabulary.

Use when changing harness gates, routing rules, skills, or automation behavior and benchmark/eval evidence is needed to avoid slowing normal coding performance. Do not use for ordinary app code changes.

Use when a request is early, ambiguous, or lacks a codebase-backed shared language; interview the user to expose assumptions before planning.

Use when a codebase, docs, ADRs, tests, prior reviews, or team traces exist; align new work with existing shared language before implementation.

Use first for normal or high-risk natural-language work to choose task mode, tier, grill path, spec layers, team-rule scan, and required gates before implementation.

Use before claiming normal+ completion when accepted REQ statements must be mapped to executed tests, guardrails, smoke checks, runtime/UI evidence, or named manual review.

Use when behavior changes need L1/L2/L3 intent specification, EARS requirements, feature archetype prompts, latency contracts, or rollback/idempotency decisions.

Use when existing codebase, docs, tests, reviews, or PR traces reveal repeated team conventions that might become reviewed rules.

Use when UI work includes forms, navigation, modals, destructive actions, keyboard/focus behavior, color contrast, aria labels, or other accessibility-sensitive flows. Do not use as the only proof of overall UX quality.

Use when a local or deployed browser route must be opened, reproduced, interacted with, screenshot, or checked for console/runtime errors. Do not use when static tests fully cover the change.

Use when changing harness gates, routing rules, skills, or automation behavior and benchmark/eval evidence is needed to avoid slowing normal coding performance. Do not use for ordinary app code changes.

Use when completing normal or high-risk code work to collect command, test, build, typecheck, runtime, review, and residual-risk evidence before claiming completion. Do not use for trivial one-line edits unless validation is ambiguous.

Use for Hermes-style self-improvement proposals based on repeated failures, skipped gates, false positives, or review findings. Do not use to silently modify active AGENTS.md, hooks, or skills.

Use when rendered UI behavior, layout, interaction states, responsive behavior, accessibility, customer language, or user-facing state changes. Do not use for pure backend, docs, or non-rendered refactors.

Use for independent review of UI screenshots, layout maps, CSS/static frontend inventories, information density, customer language, responsive quality, and destructive-action clarity. Do not use as a substitute for tests or implementation review.

Use when writing, reviewing, or refactoring code to avoid hidden assumptions, overengineering, broad unrelated edits, and unverifiable success criteria.

Use when a task changes UI structure, visual design, interaction states, accessibility, responsive behavior, design systems, landing pages, dashboards, mobile screens, forms, charts, or frontend polish.

Use when reporting completion for normal or high-risk work. Verifies validation, review, documentation, residual risk, and follow-up automation needs.

Use when reviewing a plan, PRD, implementation, diff, or completion claim. Focuses on correctness, regressions, missing tests, and risk rather than style.

Use when a user gives a new product, coding, review, debugging, or planning task in this Codex harness. Classifies tier, decides whether subagents/native features are useful, and fixes the initial contract before work starts.

Use when making normal or high-risk code changes that need an implementation plan. Creates a concise plan with ownership, validation, rollback, and subagent dispatch criteria.

Use when product strategy, PRD creation, feature prioritization, opportunity analysis, acceptance criteria, or pre-mortem risk analysis is needed before implementation.

Use when a repeated failure, repeated success pattern, or recurring workflow should become a proposed Codex skill after audit, not automatic self-modification.

Pre-landing code review. Finds bugs that pass CI but break in prod. Runs after implementation and before ship. Maps to Gate ② (mutual review) for code artifacts.

Use when something broke — a bug surfaced, a test regressed, a deploy misbehaved, or a user reported weird behavior. Enforces root-cause investigation BEFORE any fix. Separates proximate cause (what broke) from root cause (why it was possible) from systemic cause (why this class of bug exists in this codebase).

Use at the START of a new product idea or feature. Reframes user's raw request into a real problem worth solving before any code is written. Inspired by YC Office Hours. Returns a focused problem statement + go/no-go.

Use after PRD is ready but before implementation. Locks architecture, data flow, edge cases, and test plan so scope cannot silently creep during build.

Use after /office-hours when the problem is defined but the plan and scope are not yet ready. Judges readiness through discovery, product, and engineering lenses. Returns a readiness verdict (ready / revise / park) + concrete gaps to close before implementation starts.

Use for UI-affecting changes before /ship. Drives an actual browser through the user flow to reproduce what the user will see — clicks, forms, error states, loading spinners. Finds the bugs that unit tests can't catch (DOM event ordering, CSS breakage, auth session edge cases, race conditions in UI state).

Final pre-release gate. Ensures tests green, docs synced, changelog updated, then pushes and opens PR. One command for the "I'm done, let's ship it" moment.

Use once plan is locked and implementation needs to start. Guides a disciplined build loop with scope freeze, milestone-sized slices, validation after each slice, and document sync. Replaces the "start coding and see what happens" failure mode with a controlled march through the plan.