Route complex work with ak:agentkit

Use ak:agentkit when a task spans several steps or domains, the correct installed Skill is unclear, or the work needs a short sequence with risk-matched verification. This Skill is a router: it classifies, selects, and dispatches. The routed Skills and Agents own the domain work.

Choose ak:agentkit only when routing adds value

Use ak:agentkit when

• The request is ambiguous, multi-step, multi-domain, or high-risk.
• Several installed Skills could apply and their order matters.
• You need to decide whether a fresh-context specialist or independent reviewer would improve the result.
• You want a visible route record and completion gates before execution.

Skip ak:agentkit when

• You named the Skill to use. Invoke that Skill directly.
• One installed Skill obviously owns a single-step task.
• The request is pure conversation, opinion, or a factual question.
• You need headless jobs across CLIs or worktrees, or coordinated multi-session teams. Those belong to separate Engineer capabilities when installed.
• The capability is not installed. Do not route to an imagined Skill or Agent.

Prepare the installed inventory

Before you start:

• Complete Onboarding, install Marketing Kit, and start a fresh session after installation.
• State the final deliverable, constraints, non-goals, acceptance criteria, and authority boundary.
• Inventory Skills and Agents actually visible to the active runtime. Installed descriptions, not remembered catalogs, are the routing authority.
• Note the highest-risk link: user-visible, credential, data, money, destructive, public-contract, deployment, or mass-audience effects increase the verification requirement.
• Decide whether the request is route-only or authorizes execution after the route is selected.

See Runtime adapters before assuming runtime parity.

Ask for a route before work begins

/ak:agentkit Route a launch-readiness audit across positioning, landing-page copy, analytics instrumentation, and privacy review. Show the shortest installed Skill chain, risk gates, owners, and evidence handoffs. Do not edit, contact anyone, publish, deploy, or spend until I approve the route

/ak:agentkit Route a launch-readiness audit across positioning, landing-page copy, analytics instrumentation, and privacy review. Show the shortest installed Skill chain, risk gates, owners, and evidence handoffs. Do not edit, contact anyone, publish, deploy, or spend until I approve the route

$ak:agentkit Route a launch-readiness audit across positioning, landing-page copy, analytics instrumentation, and privacy review. Show the shortest installed Skill chain, risk gates, owners, and evidence handoffs. Do not edit, contact anyone, publish, deploy, or spend until I approve the route

Follow the routing stages

1. Apply the proportionality gate. Use the named or obvious Skill directly; answer trivial questions without routing ceremony.
2. Classify the work. Record one workflow class, size, risk, and domain count, choosing the class of the final deliverable.
3. Inventory installed capability. Confirm every candidate Skill and Agent exists in the current runtime. Work inline or report a gap when it does not.
4. Build the shortest chain. Select one primary owner per intent and use a subset of understand → decide → execute → verify → deliver. Each link needs named entry and exit evidence.
5. Delegate only at a trigger. Use specialists for noisy discovery, independent work, testing, high-risk review, or repeated failure. Parallel work requires disjoint ownership.
6. Enforce risk gates and report. Low risk uses executor checks; elevated risk adds verification and self-review; high risk adds independent review and explicit confirmation before the irreversible action.

Longer chains should pass artifact paths and short reports rather than lossy prose summaries. A failed link does not advance: diagnose or rescope, then resume at that link. Two consecutive failures on the same link should stop with the smallest missing input.

Keep routing separate from authority

A subagent receives only its delegation contract. Every dispatch should name one outcome, files or sources to read, allowed modifications, acceptance criteria, constraints, report location, and completion status. The controlling session retains conflict resolution, approvals, and final acceptance.

Verify the routing evidence

A complete route should provide:

• Route: <class> | size: <size> | risk: <risk> | domains: <n>.
• The installed Skill or Agent that owns each link.
• Entry evidence, output artifact, and exit check for each link.
• Delegation timing and disjoint ownership where Agents are used.
• Verification and reviewer gates matched to the highest-risk effect.
• A record of links run, artifacts produced, checks completed, gaps, and actions still waiting for approval.

Troubleshoot and understand limits

The router only knows capabilities discoverable in the active installation. It does not guarantee that every runtime exposes equivalent Agent tools, Hooks, or parallel behavior. Continue with Marketing Agents for the packaged role boundaries and Marketing Kit overview for the human-controlled workflow.