AI Skills — Model Routing
Stop Marrying One Model. Build a Router Instead.
Most teams run everything through one AI model. New model drops? They switch entirely. That is not a strategy. This skill gives you a structured comparison test and a routing table you maintain over time, so each task goes to the model that actually handles it best. Setup takes about thirty minutes.
The Problem
Two of the most capable models available right now, GPT-6 Astra and Claude Opus 5.5, take completely different approaches to the same prompt.
That is not a flaw. It is structural. Different architectures, different training, different instruction-following behaviors. The model that reasons through a complex plan is not always the model that writes clean prose. The one that handles structured data extraction well is not always the one you want for nuanced judgment calls.
Most teams do not discover this. They pick one model, usually whatever was newest when they started, and route everything through it. When it falls short on a task, they assume AI cannot handle that task. The real answer is usually that they used the wrong model.
The correct question is not "which model is better?" The correct question is "which model handles this specific type of task best for my use cases?" That answer is different for every business. It changes as models update. You only get it from direct comparison.
This skill gives you the comparison framework to build your own routing table. Once you have it, you check it before setting up any new workflow.
Copy: Model Routing Test Prompt
Run this prompt on each model you want to compare. It asks the model to evaluate itself honestly across your actual task types. You compare outputs side by side, note where the approaches diverge, and build your routing table from the patterns.
Replace the bracketed placeholders with your actual task types. Run it on both models. Compare.
Copy this.
You are helping me build a model routing table for my business. I need to understand how you handle different types of tasks so I can match work to the right model. For each task I list below, give me: 1. Your approach — how you would handle this specifically 2. Your output format — what the result looks like (list, table, prose, structured data, etc.) 3. Your honest limitations — where your approach is most likely to fall short on this task type 4. A confidence level — high / medium / low — based on how reliably you produce quality output here Be specific and honest. I am comparing you against another model, so self-promotion does not help me. My task types: 1. [TASK TYPE 1 — e.g. "Rewrite cold outreach emails with a specific tone"] 2. [TASK TYPE 2 — e.g. "Debug scripts and explain what broke"] 3. [TASK TYPE 3 — e.g. "Extract structured data from unformatted text"] 4. [TASK TYPE 4 — e.g. "Write positioning copy for service offers"] 5. [TASK TYPE 5 — e.g. "Plan and sequence multi-step workflows"] After evaluating each task, give me a one-line routing recommendation: "Use me for this" or "Another model may do better here — here is why."
How to Use It
Run the prompt on GPT-6 Astra. Run it on Claude Opus 5.5. Put both outputs side by side in a document. Look for where they disagree on their own limitations — those are the task types where the routing decision matters most.
What You Get Back
01. Side-by-side model self-assessments for each task type you specified. Both models evaluated against the same criteria.
02. Divergence points where the two models give different self-assessments for the same task type. These are your highest-value routing decisions. If both models rate themselves "high" on the same task, pick either. If they diverge, run a real test case before deciding.
03. A draft routing table you can refine after one week of actual usage. The model's self-assessment is a starting point. Your experience using both on real work is the final answer.
04. A decision rule for new workflows. Before you set up any new automation, check the routing table. If the task type is already on the list, the model is decided. If not, run a quick comparison test before committing.
05. A reusable comparison template. The same prompt works as either model gets updated. Keep the template, update the task list as your workflows evolve.
Honest Limits
Model self-assessment is not perfect. Both models can be overconfident on tasks where they actually struggle. Use the self-assessment to identify where to run real test cases, not as the final verdict.
Routing tables go stale. Both GPT-6 Astra and Claude Opus 5.5 receive ongoing updates. A routing decision that was correct three months ago may not be correct now. Run the comparison test whenever a model ships a significant update, especially to tool use, reasoning, or instruction-following.
Some tasks are genuinely model-neutral. Clean extraction from well-formatted data, basic summarization, simple classifications: either model handles these reliably. Do not over-route. Save the routing discipline for the tasks where model choice actually changes the output quality.
Most teams will not build this table. They will keep defaulting to one model and blaming AI when the wrong one fails on a task the right one handles well.
Run the comparison this week on your top five task types. Build a one-page routing table. That document will save you more debugging time than any prompt engineering tutorial.