메뉴
Recommends which framework agent (or multi-agent workflow) should handle a task; use when the user asks which agent to use, how to delegate a piece of work, or how to sequence agents for a multi-domain task.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Verify the framework's toolchain (git, bash, jq, gh, yq, PowerShell 7+, Node.js/npx, uv/uvx, shellcheck, Claude Code CLI) is installed and working — use when setting up the framework, after system updates, or when hooks, validators, MCP servers, or the executor agents fail with "command not found".
Guide for adding a new real Claude Code hook to this framework as a .ps1/.sh pair (registration in hooks/hooks.json as a shell-form chain, tests in both suites, validation) — use when asked to create, generate, or extend a hook or quality gate.
Diagnose and fix agent routing, loading, and configuration issues by inspecting agents/*.md frontmatter, the claude.json registry, hook registration, and validator output — use when an agent is "not found", tasks route to the wrong specialist, models mismatch, hooks do not fire, or validate-consistency.sh fails.
Git branching, commit, PR, conflict-resolution, and release guidance integrated with the framework's peer-review final gate — use when planning branches, writing commit messages, preparing or opening PRs, resolving merge conflicts, cutting releases, or when the Stop gate blocks a session over unreviewed committed work.
Iteratively improve a code change by having the matching specialist agent score the diff 0-100 against an explicit rubric, rewriting the weakest parts, and repeating until the score plateaus. Use after the quality bar is green (build/lint/format/tests clean) but before the review gates — e.g. polishing an implementation or refactor, or when asked to "clean up" code without concrete criteria. Code-side sibling of self-scoring-loop.
Iteratively improve a non-code deliverable (spec, plan, document, design) by writing an explicit rubric, scoring the work 0-100 against it, rewriting the weakest parts, and repeating until the score plateaus. Use when a deliverable has a quality bar but no test suite — e.g. before handing over a spec, proposal, architecture doc, or report.
| name | agent-routing-advisor |
| description | Recommends which framework agent (or multi-agent workflow) should handle a task; use when the user asks which agent to use, how to delegate a piece of work, or how to sequence agents for a multi-domain task. |
Analyze a task and recommend the right agent(s) from this framework's roster, plus a delegation prompt and workflow sequence. Recommendations must reflect the registry (claude.json), never memory or legacy docs.
.sub_agents; each agent's definition is agents/<name>.md; canonical categories are in .agent_categories.| Category | Agents | Route here when |
|---|---|---|
| language_experts | rust-expert, csharp-expert, go-expert, java-expert, python-expert, typescript-expert, mql-trading-dev | Implementation in a specific language/stack, including backend/API work |
| automation_experts | bash-expert, powershell-expert | Long, output-heavy shell runs (selective policy — short commands run inline), shell scripting, OS automation |
| domain_specialists | database-specialist, frontend-specialist, security-specialist, uiux-specialist | Schema/query work, UI components, security audits, UX/accessibility |
| infrastructure_operations | devops-orchestrator | CI/CD, containers, IaC, deployment |
| architecture_planning | system-architect, product-owner | System design, technical decisions; requirements, user stories |
| quality_analysis | comprehensive-analyst, code-review-gatekeeper, peer-review-critic | Investigation/profiling; code review; final diff-scoped peer review |
| documentation | technical-docs-writer | Guides, API docs, developer documentation |
Match dominant signals in the task description:
Decision order: (0) the task is purely running/summarizing commands → bash-expert or powershell-expert; (1) single language implementation → that language expert; (2) clearly one specialized domain → that specialist; (3) spans domains → multi-agent workflow; (4) ambiguous or investigative → comprehensive-analyst first.
Every workflow that produces or changes code ends with, in this order:
Never place peer-review-critic mid-workflow, and never end a coding workflow without it.
Feature development product-owner (stories, acceptance criteria) → system-architect (design) → [language expert] (implementation + tests) → technical-docs-writer (docs) → code-review-gatekeeper → peer-review-critic
Security-critical feature security-specialist (requirements, threat model) → [language expert] (secure implementation) → security-specialist (audit) → code-review-gatekeeper → peer-review-critic
Full-stack feature system-architect (design) → database-specialist (schema, migrations) → [language expert] (API/backend) → frontend-specialist (UI) → uiux-specialist (UX/accessibility validation) → code-review-gatekeeper → peer-review-critic
Performance optimization comprehensive-analyst (profile, identify bottlenecks) → [language expert] (optimizations) → comprehensive-analyst (verify improvement) → code-review-gatekeeper → peer-review-critic
Skip phases that don't apply (e.g., no database-specialist if no schema change), but never skip the two closing review gates when code changed.
If the task is underspecified (e.g., "optimize performance", "improve the app"):
When handing the recommendation to /agentic-framework:delegate or a Task invocation, structure the prompt as:
Present each recommendation as:
Do not include invented time estimates, pass rates, or cost figures — only what can be derived from the task and the registry.