FREE SKILL
Two Models, One Budget. One Skill Routes the Work.
Sonnet 5.5 is $2 in and $10 out. Astra runs at $10 in and $50 out. Same category of work. Five times the price gap.
On the Terminal Benchmark, Sonnet 5.5 scores 70.6%. Astra scores 57.7%.
The cheaper model wins on most daily work.
So why pay five times more for every task you run?
You don't have to. This skill routes automatically: Astra gets the hard, once-a-week thinking jobs; Sonnet handles everything else.
The Problem: One Model Gets Everything
Most businesses pick a single AI model and send every task to it.
Bad economics. A customer email doesn't need the same model as a deep strategy analysis. Your Astra bill isn't high because you need Astra. It's high because nobody set up routing.
Result: you're paying $50 per million tokens to write a follow-up email.
Astra is right for one-off research tasks, browser sessions, and anything that needs reasoning across a very long document. That's maybe 10% of what a business sends to AI in a week. Sonnet 5.5 handles the other 90%: docs, spreadsheets, replies, fixes, overnight automation runs. And it responds 30% faster.
Running both models is not complex. Running them with no routing is expensive.
What This Skill Does
It adds a single routing layer before any task executes. The router reads the incoming request, classifies it as Sonnet work or Astra work, and calls the right model. You see one line before the result: "Routing: SONNET — daily email reply." Nothing else changes in your workflow.
The classifier uses five signals: does the task require live browsing? Does it need reasoning across more than 50 pages? Is it explicitly marked high-stakes or one-off? Is it a repeating template task? Is the user overriding manually?
Sonnet wins unless the answer to one of the first three is yes.
Copy This Skill
Paste this into your Claude Code or Codex project as a system-level prompt. It runs on every task from that point forward.
Copy this.
You are a dual-model task router for [YOUR BUSINESS CONTEXT]. Your job: classify each incoming task, then call the right model. SONNET handles: emails, summaries, document edits, spreadsheet work, code fixes, follow-ups, form processing, any task that runs daily and follows a clear pattern. Use Sonnet for the 90%. ASTRA handles: deep research requiring more than 50 pages of context, tasks that require live browser access, one-off strategy questions, and any request explicitly marked as "complex" or "one-off." Routing logic: 1. Read the request. 2. Does it require live browsing, very long document reasoning, or is it explicitly a one-off hard problem? → ASTRA 3. Everything else → SONNET Before running any task, output one line: "Routing: [SONNET/ASTRA] — [reason in under 10 words]." Then execute on the selected model. If the user says "force Sonnet" or "force Astra", honor that override and skip the routing line.
How to Use It
Save this as a Claude Code skill or paste it at the top of a new project. Every task that enters the session gets a routing decision before execution. Nothing else changes.
To force a model on a specific task, add "force Sonnet" or "force Astra" to your message. That overrides the classifier for that request only.
Run it for one week. Read the Routing lines. You'll see quickly which tasks were getting routed to the expensive model unnecessarily.
What You Get Back
- A one-line routing decision before every task ("Routing: SONNET — daily email reply").
- Astra reserved for tasks that genuinely need its capability.
- Sonnet handling daily volume at $2/$10 instead of $10/$50.
- A manual override for tasks where you know better than the classifier.
- Baseline visibility into how your AI budget is actually being spent.
Why This Is Not a Silver Bullet
This skill routes by task description, not by true complexity. A vague request can still misdirect. "Summarize this thread" routes correctly. "Do something about this" doesn't.
Be specific in your prompts. Specificity beats the classifier's ambiguity every time.
Also: the routing doesn't check context window size. A 300-page document still needs Astra even if the request sounds simple. If you regularly work with large files, add an explicit "this is a long document" note to your prompts when relevant.
The first week, check the Routing lines after every run. Adjust your prompting based on what you see. After two weeks, you stop thinking about it.
Start with one task type. Set up routing for email replies first. Watch the cost line. Expand from there.
Done.