Task Router

Task Router

Pick the execution shape first, the model second. Wrong shape is the expensive mistake — a fan-out for a typo burns tokens and wall-clock; an inline guess at an architecture change burns an afternoon.

This skill is the authorization to delegate. The standing rule in this environment is don't use the Agent tool or Workflows unless the user, a CLAUDE.md, or a skill asks for it. This skill asks for it — but only at the tier it selects. Tier 0 still means no agent.

When to use

• Before spawning any subagent or workflow.
• When you can't yet name the files a request touches.
• When a request spans multiple repos, domains, or review axes.
• When choosing which model to hand a delegated task.
• When the user asks how to approach, size, delegate, or staff a piece of work.

Don't use it for a request whose target and scope are already obvious — that is Tier 0 by inspection, and routing it is ceremony.

1. Model policy — read this first

This is a hard constraint, not a preference. It comes before any efficiency argument.

🚫 The excluded tier

Most setups have at least one model tier that is off-limits — it bills outside the plan, it isn't approved, or it costs more than any task here justifies. Name that tier explicitly and never pass it, because tool APIs accept a model name silently: nothing warns you that you just left your billing plan. If a task seems to want the excluded tier, it wants the highest permitted one instead.

Worked example — the setup this skill was written against. A Claude Max subscription where fable (Fable 5.1, claude-fable-5-1) bills as direct API tokens rather than against the subscription, so it is excluded outright. Substitute your own excluded tier; the routing logic below doesn't change.

The tiers

Named for Claude models below. On another provider, map the three roles — mechanical, default, hard — onto the equivalent tiers and keep the table's logic.

Documentation and basic research are sonnet, not haiku. They read as cheap but they are judgment work: what to include, what's stale, what the source actually says. Haiku is only for passes where the answer is mechanical.

Omitting model inherits the parent's model, which is fine and often correct. When genuinely unsure between two tiers, take sonnet.

2. Score the request

Answer these before choosing a tier. Each is countable, not a vibe.

Risk overrides size. A one-line change that deploys to production is not Tier 0. Small and irreversible means: do the work, then stop and confirm before the irreversible step.

Ambiguity is not complexity. Q5 firing means the request is underspecified — more agents will not fix that. One clarifying question beats a confident fan-out.

3. Pick the tier

Tier 0 — Inline (no agent)

When: the files are known and there are three or fewer, or the answer is recall. Do: just do it. Read, edit, answer. This is the default. Most requests stop here.

Tier 1 — One subagent

When: scoped to one domain, the target is known, and the result is independently verifiable. Do: one Agent call. Agent by domain (§5), model by §1. Why bother: it keeps a large read out of the main context. Delegate when the reading is big, not merely when the writing is.

Tier 2 — Recon or plan, then act

When: you cannot name the files yet, or the change touches shared code where the real fix is upstream of the reported symptom. Do: Explore (locate) or Plan (design) → then execute at Tier 0/1 with what came back. Never skip to Tier 1 here — a subagent that has to find its own target will guess.

Tier 3 — Fan-out

When: three or more genuinely independent dimensions and the user opted in. Do: parallel Agent calls in a single message, or Workflow if the user asked for one. Gate: if the dimensions share state or must run in order, it is Tier 2, not Tier 3.

Escalating a tier costs one more call; de-escalating means throwing away finished work. Start one tier low.

4. Report the decision

State the routing decision in one line before acting, so it is cheap to override:

Tier 1 · security-auditor · sonnet — scoped to one repo, findings are verifiable.

Then act. Do not narrate the scoring.

5. Agent selection

Check the project's .claude/agents/ first — a project-scoped agent knows the repo. Fall back to user scope. If nothing fits, general-purpose.

6. Calibration

These are the reference points. If a routing decision disagrees with the nearest row, re-check §2 before overriding it.

Anti-patterns

• Passing the excluded model tier. Ever, for anything — nothing warns you that you just left your billing plan.
• Sending documentation or research to haiku because it "looks cheap." Both are judgment work — they are sonnet.
• Fanning out to look thorough. Three agents on one dimension is one agent plus noise.
• Tier 1 with an unknown target. The subagent will guess at what to change. Recon first.
• Routing an ambiguous request. If two readings give different work, ask one question.
• Treating a small diff as low risk. Reversibility, not line count, sets the risk gate.
• Narrating the scoring. One line, then act.
• Opus for bulk mechanical work. That is what the tiers are for.