# Skill Router > Choose one next engineering skill when the user explicitly asks which skill to use or invokes $skill-router. Return that skill or none, then stop without starting it. - Skill: `stevennitesh/skill-router` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add stevennitesh/skill-router` - Raw SKILL.md: https://api.skillmd.com/api/skills/stevennitesh/skill-router/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: stevennitesh (https://skillmd.com/u/stevennitesh) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/stevennitesh/skill-router --- # Skill Router **Route:** return exactly one next skill or a truthful `none`, then stop. Downstream skills own their procedures, artifacts, mutations, proof, and completion. ## Spine 1. **Inspect.** Use the user's stated situation and visible repo state. Inspect only a fact that could change the route. 2. **Clarify.** If two routes remain plausible, ask one highest-leverage question and wait. 3. **Route.** Choose one exact route below. After a candidate wins, inspect only its applicable installed `docs/agents/*` contract when compatibility could change the route; never load Repo Bootstrap guidance. If that contract is missing, incompatible, or outdated, route to `$repo-bootstrap` instead. If no available skill satisfies its entry contract, select `none`; never substitute the nearest or weakest route. 4. **Stop.** Return `Skill: ` and `Reason: `. Add `Precondition: ` only when the user must satisfy one before starting the route. The user starts any selected skill; downstream work remains unstarted. `none` is a terminal abstention, not a recommendation. Use it only when current facts make every available route ineligible, not instead of one allowed clarification or a required `$repo-bootstrap` route. ## Route Map ### Shape | Situation | Route | | --- | --- | | One current-user-owned repo-backed decision needs live grilling and domain reconciliation as answers settle | `$grill-with-docs` | | The current user owns one bounded decision needing conversation-only stress-testing | `$grilling` | | A bounded destination has interdependent decisions, a non-conversational resolver, and needs multi-session tracker sequencing | `$wayfinder` | | One external stakeholder holds missing knowledge and needs an async discovery questionnaire | `$to-questionnaire` | | One bounded source-answerable question needs cited evidence | `$research` | | One design question needs disposable runnable evidence | `$prototype` | | A fresh context that can read the same work root needs a verified local pickup | `$handoff` | **Unknown-owner tie-breaker:** route a source-answerable fact to `$research`, a runnable design choice to `$prototype`, an external-stakeholder gap to `$to-questionnaire`, and a current-user decision to `$grilling` or `$grill-with-docs` when domain meaning may change during the conversation. Route to `$wayfinder` only after the destination is bounded and several interdependent decisions or prerequisites, including at least one non-conversational resolver, need a tracker-backed multi-session route. ### Build | Situation | Route | | --- | --- | | Settled source needs a durable parent decision contract before ticket slicing | `$to-spec` | | A verified parent specification or equivalent settled bounded source needs a dependency-ordered implementation ticket graph and actionable frontier | `$to-tickets` | | One bounded settled implementation is selected directly or as a Ready-for-agent item | `$implement` | | One explicitly requested bounded campaign must repeatedly improve one frozen measurable runtime, resource, cost, capacity, or product outcome against a settled target | `$hillclimb` | | One explicitly requested bounded procedure has at least two dependent settled stages only the current human can perform, and scripting materially reduces secret exposure or configuration mistakes | `$wizard` | | One explicit fixed delivery set has at least two accepted implementation items and a non-empty ready frontier | `$parallel-implement` | Route one selected item to `$implement`; route an explicitly requested fixed set of at least two accepted items with a non-empty ready frontier to `$parallel-implement`. Parallel Implement checks dependencies, ownership, and write effects before concurrent dispatch. Route one settled optimization to `$implement`, a sustained runtime, resource, cost, capacity, or product target to `$hillclimb`, measured code-shape or maintainability reduction to `$simplify-code`, and unexplained slowness or regression to `$diagnosing-bugs`. Route agent-executable setup or maintained automation to `$implement`, one manual action to no skill, and at least two dependent settled human-only stages that benefit materially from an interactive script to `$wizard`. An unresolved decision, source fact, or stakeholder answer remains with its shaping or evidence owner. ### Incoming Work And Quality | Situation | Route | | --- | --- | | Raw tracker issues or configured external PR/MR requests need sorting and readiness verification | `$triage` | | A hard, intermittent, performance, environment-only, production-only, or causally ambiguous failure needs dedicated investigation | `$diagnosing-bugs` | | The user explicitly requests TDD, test-first work, or RED-GREEN-REFACTOR, or applicable repository policy requires TDD, and one bounded behavior and independent oracle are settled | `$tdd` | | An active merge, rebase, cherry-pick, or revert is conflicted, or an index is unmerged | `$resolving-merge-conflicts` | | A branch, WIP, staged, since-X diff, PR, release candidate, or supported-risk candidate needs read-only judgment | `$change-review` | | The user explicitly requests high-assurance, heavy, or final review of one fixed complete code candidate | `$high-assurance-review` | | A repository needs a whole-system HTML map, one user-selected subsystem audit, or one user-selected audit-candidate analysis | `$audit-codebase` | | One user-selected existing-code target has accepted behavior and needs behavior-preserving simplification | `$simplify-code` | **Existing-code tie-breaker:** route repository-wide discovery or baseline judgment to `$audit-codebase`; one user-selected target with accepted behavior to `$simplify-code`; and one module, interface, or seam decision to `$codebase-design`. Route a selected ready item to `$implement`, which invokes TDD only under an explicit user or repository-policy requirement. Route one standalone explicitly test-first behavior to `$tdd`, ordinary test, integration-test, regression-test, or coverage work to `$implement`, uncertain broken behavior that needs dedicated investigation to `$diagnosing-bugs`, and an ordinary existing diff needing judgment to `$change-review`. Route to `$high-assurance-review` only for an explicit heavy-review request; candidate size, PR presence, release status, novelty, or risk does not qualify. **Conflict tie-breaker:** route an active unresolved operation or unmerged index to `$resolving-merge-conflicts`; an already-resolved candidate to review; a routine post-operation repair to `$implement`; and a hard post-operation failure that needs dedicated investigation to `$diagnosing-bugs`. ### Design And Pack Maintenance | Situation | Route | | --- | --- | | Decide the architecture of one bounded module or interface | `$codebase-design` | | Resolve or capture domain terms, invariants, context boundaries, or relationships; assess or record an already-settled ADR candidate | `$domain-modeling` | | Create, edit, or audit instructions agents consume | `$writing-for-agents` | | Persistent coding-agent context needs memory admission, reflection, audit, or approved pruning | `$context-hygiene` | `$domain-modeling` and `$codebase-design` are shared disciplines. Route unresolved code shape, interfaces, or module ownership to Codebase Design; route project-specific semantic truth or an already-settled ADR candidate to Domain Modeling. Otherwise let the owning workflow proceed directly. **Handoff / compact:** `$handoff` starts a fresh same-root context; `/compact` continues the current conversation.