model-hierarchy

Overview of model-hierarchy skill

What model-hierarchy does

The model-hierarchy skill helps an agent route work to the cheapest model that can handle it. It is built for situations where you want better cost control without losing quality on routine tasks. If your workflow spends premium tokens on file reads, status checks, formatting, or simple lookups, this skill gives you a practical model-hierarchy guide instead of relying on intuition alone.

Who should install it

Install model-hierarchy if you run an agent workflow that spawns sub-agents, switches models often, or pays for many small tasks. It is especially useful for Workflow Automation, Claude Code-style setups, and any environment where the wrong model choice quietly inflates spend. It is less useful if you already have strict routing logic in code or you rarely change models.

What makes it different

The skill is not just “use a cheaper model.” It encodes a simple decision rule: routine tasks go low-tier, moderate tasks stay mid-tier, and only genuinely hard problems justify premium reasoning. That makes model-hierarchy more actionable than a generic prompt because it gives the agent a repeatable classification habit and a clear default for sub-agent work.

How to Use model-hierarchy skill

Install model-hierarchy

The repository is designed as a skill you copy into your agent’s skill directory or prompt context. For OpenClaw, the repo README shows copying SKILL.md into the skills path and restarting the gateway. For Claude Code / Codex-style systems, the practical install is to paste the routing rules into CLAUDE.md or your project instructions. If you are evaluating model-hierarchy install, check whether your agent reads skills from files, global instructions, or a repo-local config.

Start with the right input

model-hierarchy usage works best when you tell the agent three things: the task type, the expected output, and whether the task is part of a larger workflow. Weak input: “help me with this repo.” Stronger input: “Classify this as routine or moderate, then choose the cheapest model that can safely read config.json, summarize the result, and report the risk if the classification is wrong.” That gives the skill enough context to route correctly.

Read these files first

Begin with SKILL.md for the routing rules, then inspect README.md for installation patterns and examples/claude-code.md or examples/openclaw.md for platform-specific usage. If you want to understand edge behavior, tests/scenarios.json is useful because it reveals how the skill classifies routine versus moderate tasks. This is the fastest path to understanding the model-hierarchy skill without reading every line of the repo.

Use it in a workflow

A practical model-hierarchy workflow is: classify the task, decide whether it is routine/moderate/complex, then select the cheapest acceptable model before execution. For sub-agents, default cheap unless the job needs deep reasoning or vision. Be explicit when the task includes image input, chart reading, or other non-text work, because text-only models should not be used there. That boundary matters more than token cost.

model-hierarchy skill FAQ

Is model-hierarchy only for OpenClaw?

No. OpenClaw is one supported integration pattern, but the skill is also relevant to Claude Code, Codex, and other agent stacks that let you define routing behavior in instructions. If your system can follow a model selection policy, model-hierarchy can usually fit.

How is this different from a normal prompt?

A normal prompt asks for one-off behavior. The model-hierarchy skill gives you a reusable routing rule that an agent can apply before every task. That makes it better for repeated operations, background agents, and cost-sensitive workflows where model choice is part of the job.

Is it beginner-friendly?

Yes, if you can distinguish routine, moderate, and complex tasks. The skill is simpler than a full policy engine, but you still need to be honest about task difficulty. If you under-classify hard debugging or vision work as routine, the cost savings disappear when the model fails and you have to rerun.

When should I not use it?

Do not use model-hierarchy as a blanket downgrade strategy for everything. If the work requires deep debugging, architecture decisions, security review, or multimodal input, the cheapest model is usually the wrong choice. It is also a poor fit if your organization already enforces model selection in code with strong guardrails.

How to Improve model-hierarchy skill

Give stronger task labels

The fastest way to improve model-hierarchy results is to state the task category up front. Good inputs name the action and the expected complexity: “routine file lookup,” “moderate code draft,” or “complex debugging with prior failure.” That reduces guesswork and helps the agent apply the right tier on the first pass.

Describe constraints that affect routing

Model choice changes when you mention context size, multimodal input, or failure tolerance. For example: “This is a text-only summarization task from a 200-line log” or “This requires screenshot analysis, so avoid text-only models.” Those details matter because they expose misfit cases the skill should not optimize away.

Iterate after the first pass

If the first output feels over-engineered, ask the agent to reclassify the task and explain why it chose that tier. If it feels too weak, request an upgrade and identify the missing signals: cross-file reasoning, ambiguity, or prior failure. The model-hierarchy guide works best when you use it as a routing check, not a one-shot verdict.

Watch for common failure modes

The main failure mode is treating “simple-looking” tasks as routine when they actually hide dependencies, edge cases, or vision needs. Another is copying the skill into a workflow without telling the agent where to find the policy or when to override it. To improve model-hierarchy for Workflow Automation, keep the routing rule close to the task source and make the escalation path explicit.