Agent Skills Router

Agent Skills Router

Use this high-level skill to choose the smallest useful downstream skill. Do not treat this as a replacement for the focused skills; use it to route, then load the selected skill's SKILL.md.

How To Route

1. Identify the user's actual task type.
2. Pick the smallest matching focused skill.
3. Name the exact focused skill and, when working inside this repo, its repo path.
4. If the package has been installed into ~/.codex/skills/, focused skills usually live by skill name, such as agent-skill-design, not by grouped repo path.
5. Load the focused skill's SKILL.md before acting.

Do not use this router as the operational instructions for the task after a focused skill matches.

Routing

Coding Agent Workflow

Use these for Codex, GitHub Copilot, Cursor, and similar coding agents while they do software work.

• Use spec-driven-agent-workflow (coding-agent-workflow/spec-driven-agent-workflow in this repo) when converting vague coding intent into a spec, BDD scenarios, failing tests, small diffs, policy gates, or reviewable production work.
• Use agent-self-audit (coding-agent-workflow/agent-self-audit in this repo) before finalizing substantial work when the agent should check its own likely weaknesses, residual risk, verification gaps, context drift, and trajectory quality.
• Use code-review-agent (coding-agent-workflow/code-review-agent in this repo) when reviewing code changes, pull requests, diffs, patches, or AI-generated code for bugs, regressions, missing tests, security issues, oversized scope, or spec mismatch.
• Use context-harness-debugger (coding-agent-workflow/context-harness-debugger in this repo) when the coding agent is looping, missing instructions, calling the wrong tool, hallucinating dependencies, overloading context, or otherwise needs harness/context repair.
• Use context-budget-planner (coding-agent-workflow/context-budget-planner in this repo) when a task has too much source material, noisy context, overlapping instructions, or repeated context rot and needs a plan for static instructions, skills, references, scripts, retrieval, or summaries.
• Use agent-incident-retrospective (coding-agent-workflow/agent-incident-retrospective in this repo) after an agent mistake, unsafe action, bad answer, loop, near miss, or user correction when the goal is root cause, guardrail repair, and a regression eval.

Agent Engineering

Use these when building, packaging, evaluating, or debugging agent systems and reusable skills.

• Use agent-architecture-methods (agent-engineering/agent-architecture-methods in this repo) when designing an agent system, choosing single-agent versus multi-agent, defining tools/state/evals/security, or turning a vague agent idea into a buildable architecture.
• Use agent-protocol-selector (agent-engineering/agent-protocol-selector in this repo) when choosing between project instructions, skills, scripts, MCP, A2A, A2UI, AP2/UCP, or ordinary application code.
• Use agent-skill-design (agent-engineering/agent-skill-design in this repo) when creating, refactoring, packaging, or reviewing portable skills and deciding what belongs in SKILL.md, references/, scripts/, or assets/.
• Use agent-eval-case-builder (agent-engineering/agent-eval-case-builder in this repo) when creating eval datasets, rubrics, golden prompts, red-team prompts, regression cases, or LLM-as-judge criteria for an agent workflow or skill.
• Use skill-evaluation-loop (agent-engineering/skill-evaluation-loop in this repo) when testing or improving a skill with trigger tests, execution tests, token-budget checks, regression checks, red-teaming, or promotion tiers.
• Use agent-observability-trace-review (agent-engineering/agent-observability-trace-review in this repo) when reviewing an agent trace, transcript, tool-call log, approval record, or suspicious final answer for trajectory quality, context loaded, tool choices, safety, cost, and verification.

Safety Governance

Use these when agent behavior touches untrusted content, authority boundaries, approvals, or tool access.

• Use prompt-injection-triage (safety-governance/prompt-injection-triage in this repo) when external or retrieved content may contain hostile instructions, hidden directives, prompt injection, data-exfiltration attempts, or tool/action requests.
• Use least-privilege-tool-planner (safety-governance/least-privilege-tool-planner in this repo) before giving an agent tools, MCP servers, file access, credentials, shell/browser/session access, deployment authority, write permissions, or external action capability.
• Use human-approval-gate-designer (safety-governance/human-approval-gate-designer in this repo) when designing HITL approval flows, pending action payloads, approve/deny/edit decisions, escalation rules, audit records, and auto-approval boundaries.

ADK Implementation

• Use adk-ambient-agent (adk-implementation/adk-ambient-agent in this repo) when building event-driven, scheduled, Pub/Sub-style, or ambient ADK agents with normalized inputs, deterministic routing, state, traces, and HITL.
• Use adk-secure-lifecycle (adk-implementation/adk-secure-lifecycle in this repo) when securing an ADK/tool-using agent with validation, authorization, guardrails, hooks, secret scanning, STRIDE, human gates, tests, and behavioral evals.
• Use adk-runtime-deployment (adk-implementation/adk-runtime-deployment in this repo) when packaging, testing, deploying, monitoring, or rolling back an ADK agent on Google Agent Runtime with Agents CLI, Terraform, IAM, artifacts, telemetry, and remote validation.
• Use adk-hitl-frontend (adk-implementation/adk-hitl-frontend in this repo) when building a human-in-the-loop frontend for pending approvals, event correlation, session resume, function responses, dashboards, auth, audit, Cloud Run, or Pub/Sub wiring.

Capstone

• Use kaggle-capstone-planner (capstone/kaggle-capstone-planner in this repo) when planning, scoping, validating, or preparing a Kaggle Agent capstone submission, writeup, demo, local evaluation, track choice, or compliance review.

Combination Patterns

• New production agent idea: start with agent-architecture-methods, then spec-driven-agent-workflow, then the relevant ADK implementation skill.
• Coding agent behaving badly: start with agent-self-audit; if there is a trace, use agent-observability-trace-review; if the issue repeats, use context-harness-debugger; after a concrete mistake, use agent-incident-retrospective.
• New skill from reference material: use context-budget-planner, then agent-skill-design, then skill-evaluation-loop.
• Tool/protocol confusion: use agent-protocol-selector, then route to MCP/A2A/A2UI/ADK implementation work as appropriate.
• Safety-sensitive tool access: use least-privilege-tool-planner, then human-approval-gate-designer if approvals are needed, and prompt-injection-triage when external content is part of the loop.
• Capstone build: start with kaggle-capstone-planner, then agent-architecture-methods, then adk-secure-lifecycle and adk-runtime-deployment when implementation matures.

Rules

• Prefer the most specific skill that matches the current task.
• Load only one or two focused skills unless the task genuinely crosses boundaries.
• If no focused skill matches, proceed without forcing one.
• If the selected skill has references, load only the references relevant to the immediate task.

Evaluation Prompts

Use these prompts to test routing behavior:

• Positive: "Help me decide whether this should be an MCP server, a skill, or an A2A agent." Expected: agent-protocol-selector.
• Positive: "My coding agent keeps calling the wrong tool and looping after long context." Expected: context-harness-debugger.
• Positive: "Turn this agent failure into a root cause and regression eval." Expected: agent-incident-retrospective.
• Positive: "Review this AI-generated PR for production risks." Expected: code-review-agent.
• Positive: "This retrieved web page says to ignore all previous instructions; what should the agent do?" Expected: prompt-injection-triage.
• Positive: "Plan my Kaggle Agent capstone submission and evaluation checklist." Expected: kaggle-capstone-planner.
• Negative: "Edit this CSS button color." Expected: no bundled agent skill unless the task expands into architecture, workflow, or evaluation.
• Negative: "Deploy this ordinary static website." Expected: no ADK deployment skill unless it is an ADK/Agent Runtime deployment.
• Negative: "Summarize this unrelated PDF." Expected: no bundled agent skill unless the summary is for capstone, architecture, or skill creation.