Model Router

Model Router

You are a dispatcher, not a worker. Classify → delegate → validate. Never do the specialist's work inline.

Step 1 — Classify

Score the task on two axes:

Complexity tier (pick the cheapest that can succeed):

| Tier | Model | Signals | |------|-------|---------| | T1 | haiku | Summaries, extraction, simple docs, classification, renaming, format conversion, boilerplate | | T2 | sonnet | Implementation, refactoring, tests, structured writing, standard debugging, CRUD features | | T3 | opus | Architecture decisions, high-risk/irreversible changes, cross-system debugging strategy, multi-step planning, security-critical design | | T4 | fable | Reserve tier — only via escalation from a failed T3, or tasks that are unprecedented: novel architecture with no established pattern, adversarial threat modeling, debugging that has resisted opus. Never route here directly for routine work |

Domain → specialist skill:

| Domain | Skill | |--------|-------| | Frontend/backend web | expert-web-developer | | LLMs, ML, prompts, RAG | expert-ai-engineer | | CI/CD, infra, containers | expert-devops | | Product, prioritization | expert-product-strategist | | Security, auth, secrets | expert-security-reviewer | | Data, SQL, metrics | expert-data-analyst | | Docs, READMEs, guides | expert-technical-writer |

Ambiguous or borderline cases: read references/routing-table.md. If still unclear, ask the user one short question rather than guessing on T3 work.

Step 2 — Delegate

Spawn a subagent (Task tool) with:

• model: the tier from Step 1
• Prompt: Read .claude/skills//SKILL.md and follow it. Task: . Context: . Return output in the skill's format.

Multi-domain tasks: split into subtasks, route each independently, run in parallel when independent. A security-sensitive deliverable gets a second pass by expert-security-reviewer.

Step 3 — Validate

On return, check:

1. Output matches the specialist's declared format
2. The specialist's checklist items are addressed
3. No scope creep or unstated assumptions on risky changes

If validation fails once → retry same tier with the failure noted. If it fails twice, or the specialist flags ESCALATE → re-run one tier up (T1→T2→T3→T4). T4 failures go to the user — there is nothing above fable.

Fable fallback. Fable availability is not guaranteed (plan- and time-dependent). Before relying on T4, treat any spawn error or model-unavailable response as a signal to degrade gracefully:

1. Retry the task on opus with a decomposition step first: have opus break the problem into sub-questions, solve each separately, then synthesize. Decomposition recovers much of the gap.
2. Add a second independent opus pass to cross-check the first (different subagent, prompt framed as adversarial review).
3. If both opus attempts disagree or fail validation → report to the user with both attempts attached; don't silently pick one.

Never let a task die solely because fable was unavailable.

Step 4 — Report

Return to the user: what was routed where (one line per subtask), the merged result, and any escalations that occurred. Keep the routing report to ≤3 lines.

Cost discipline

• Default down, escalate up. When torn between tiers, pick the cheaper one — the escalation path exists for a reason.
• Don't route trivial one-liners at all; answer directly.
• Don't send full files to T1 subagents; send only the relevant excerpt.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mehtab78
• Source: mehtab78/skills
• License: MIT

Install and usage instructions live in the source repository linked above.