Treating AI Models Like a Team: My Aha! Moment

Treating AI Models Like a Team (My Aha! Moment)

So here’s a scruffycat-style confession: I used to treat AI models like a box of random tools. Need something done? Grab whatever’s closest and hope for the best. But that’s not how these digital critters work. Each one’s got its own quirks, strengths, and blind spots. Once I started thinking of them as a team—each with a job title and personality—my workflow got way smoother (and a lot more fun).

Sometimes, even with the right roles, keeping track of the big picture can get tricky—especially when you’re swapping models or your chat history is longer than a cat’s tail. Here’s how I keep my studio running smooth:

The Senior Architects

At the time of writing: Claude Sonnet-4, Claude Sonnet-3.5, GPT-5 Codex (code-specialized), GPT-5 (general purpose), Gemini Pro, etc.

These are the big brains. You call them in when the problem is gnarly, ambiguous, or you need some serious creative horsepower. They’re slower and pricier, but they’ll sketch out the blueprint and catch stuff the juniors miss. They don’t just answer—they strategize.

The Offshore Engineer

Currently: GPT-4.1, Gemini Pro, and similar “mid-level” models in GitHub Copilot.

This is your reliable, cost-effective dev. Fast, practical, and gets most of the job done. But, like any remote teammate, you want to double-check the details. They’ll build the thing, but oversight matters.

The Juniors

Right now: GPT-4o, GPT-5-mini, GROK-Code-Fast (experimental)

Speedy, scrappy, and perfect for grunt work, cleanup, or wild brainstorming. Great for quick fixes or exploring ideas without burning through tokens. Just don’t expect them to architect your next big project solo.

How I Use This Team (with GitHub Copilot)

I’m wrangling these models every day in VSCode with GitHub Copilot. The architects help me design new features or untangle gnarly bugs. The engineer builds out the code, and the juniors handle refactoring, formatting, or wild idea generation. Sometimes I’ll swap in a new model if it’s in preview—always fun to see what the latest AI “intern” can do.

Why This Changed Everything

Before this, I’d toss everything at the fastest model and get annoyed when it missed the big picture. Or I’d waste my “senior architect” tokens on typo fixes. Now, I assign the right job to the right model, and my workflow is way more efficient (and my wallet’s happier).

It’s project management, but with AIs. The architects design, the engineer builds, the juniors sweep up and experiment.

Keeping the Big Picture (and Your Sanity) When Working With AI

One thing I’ve learned wrangling these digital teammates: context is everything. If you’re swapping models, jumping between tasks, or have a conversation history longer than a cat’s tail, you need a system to keep the big picture in focus. Here’s how I do it at Scruffy Cat Studio:

1. Copilot Needs Instructions!
   Spend time crafting your COPILOT_INSTRUCTIONS file. This is where you lay out how code should be written—syntax, doc format, OOP principles, and other high-level standards. Keep most of this constant across projects, but add repo/project-specific notes as needed. Don’t put architecture details here; treat it as a living doc. If the AI starts repeating mistakes, update the instructions to improve your workflow.

2. Co-write Your Architecture
   When architecting, don’t just chat with the model—co-author an ARCHITECTURE.md document. Spell out all aspects of your system clearly. This keeps everyone (human and AI) on the same page, no matter how many models you swap in and out.
   Pro tip: Always ask, “Is there anything that we are missing? Have we covered all aspects of the architecture?” Then iterate with your AI teammate until the document feels complete.

3. Plan Your Execution Steps
   During design, include a high-level execution plan. Figure out the order of implementation for your modules. This helps you and your AI team stay focused and avoid wandering off into the weeds.

4. Track Implementation Like a Pro
   When it’s time to build, focus on your implementation steps. I use GitHub Issues to track work, set milestones, and document what the model and I do in the comments. It’s like JIRA, but without the bloat. Alternatively, break your work into files in your repo—don’t rely on chat history alone. Keep records of what needs doing and how you did it!

5. Keep Everyone in Their Lane
   This is KEY! Don’t let the speedy models do deep reasoning. Assign tasks based on strengths—let the architects design, the engineer build, and the juniors handle grunt work. Otherwise, you’ll end up frustrated and chasing your own tail.

By keeping your instructions, architecture, plans, and records organized (and assigning the right jobs to the right models), you’ll maintain context and keep your AI team running smoothly—even when the conversation history gets wild.

Takeaway

If you’re juggling multiple models, try thinking of them as teammates with different job titles. It takes a little practice to “manage” them this way, but once you do, you’ll wonder why you didn’t start sooner.