You Taught an AI to Do Your Work. What Happens When You Switch Tools?

You’ve probably spent hours writing the perfect prompt—maybe dozens of them—so ChatGPT can handle your customer inquiries, or Claude can draft your social media posts. It feels great when it works. But then you read about a new AI tool that’s faster, or your team wants to move to a different platform, and you realise something uncomfortable: all that effort stays behind. You’d be starting from scratch.

That’s the exact problem a team of researchers from Microsoft, Shanghai Jiao Tong University, Tongji University, and Fudan University just tackled. They built something called SkillOpt, and their findings suggest that the “skills” you teach an AI agent—the step-by-step procedures it follows—might soon be portable. You could train an agent in one environment and move it to another, without losing the knowledge.

This isn’t just lab talk. It directly affects how much value you can get from AI tools without being locked into one vendor.

TL;DR: Microsoft researchers trained AI agent “skills” (plain-text instruction files) in one tool, then moved them to another harness like Codex or Claude Code. Results show procedural skills transfer surprisingly well—one spreadsheet skill even outperformed the target tool’s own in-house training. For Malaysian SMEs, this means your AI workflows could become portable across tools, preserving your investment in prompts and procedures.

What This Means

Let’s break this down without the jargon. An AI agent isn’t just a chatbot. It’s an assistant that can use tools—run calculations, search your files, write code, update spreadsheets. To get it to do useful work, you often need to give it a “skill”: a set of instructions, examples, and rules for completing a specific task, like “reconcile these bank transactions” or “generate a monthly report.”

Currently, these skills are usually tied to a specific AI system. The way ChatGPT handles a task is different from how Claude handles it. The underlying models think differently. So if you switch providers, your carefully crafted instructions may break.

SkillOpt is a method that creates a single, text-based skill document (literally called best_skill.md) that works across different AI models and different agent harnesses. The researchers optimise this document using an “optimiser model” that scores trial runs and suggests small edits—add a rule, delete a useless instruction, replace a step. Only the edits that improve performance are kept. The final result is one file that any compatible agent can read.

And here’s the key part: the researchers proved this file can be transferred. Train it on one model, deploy it on another. Train it inside OpenAI’s Codex, run it inside Anthropic’s Claude Code. It’s not always perfect, but it often retains most of the benefit—and sometimes exceeds it.

How This Applies to Malaysian SMEs

You might be thinking, “I’m not a software engineer. I just use ChatGPT to reply to customer emails.” Fair enough. But consider how many of your business operations depend on instructions you’ve built into a specific AI tool. Maybe you have a chatbot on your website, trained on your product catalogue and FAQ. Or you’ve set up an automated system that reads received invoices and categorises expenses. Or you use an AI assistant to draft quotations based on a specific format you designed.

All of those are “skills.” And today, they’re mostly shackled to one platform. If that platform changes its pricing, or you find a better tool for a specific job, you either start over or stay trapped. SkillOpt’s findings suggest a future where that skill document is portable—where the knowledge you’ve encoded isn’t held hostage by a vendor.

For a small business owner in Malaysia, this has three practical implications.

First, your AI investments can become assets. Right now, when you spend two weeks teaching an AI agent how to handle your e-commerce returns process, that knowledge lives inside one platform. If you switch platforms, it’s gone. The research shows that procedural skills—step-by-step workflows like “inspect the spreadsheet structure, verify formulas, then write values”—can move across models with high retention. In one test, a skill trained on a larger model retained 82% of its performance when moved to a smaller model from the same family. More impressively, a spreadsheet skill trained inside Codex achieved 81.8 points when used in Claude Code, slightly above the 80.4 points Claude Code achieved by training the same skill from scratch itself. That’s basically free transfer.

Second, you can choose the right tool for the job. Malaysian SMEs often juggle multiple platforms—WhatsApp Business, email, accounting software like SQL or AutoCount, and e-commerce marketplaces like Shopee or Lazada. If an AI skill can move across environments, you can optimise it where it’s easiest to train, then deploy it where your business actually runs. The research notes that all execution modes consume the same skill file format, which is what makes the transfer possible. That means the agent you use for your website doesn’t have to be the same one you use for your internal operations. You can pick and choose based on capability, not just lock-in.

Third, you gain auditability. One underrated finding is that the deployed skill is a text file a domain practitioner can read in minutes—not a black-box neural network. The researchers’ system records every edit it accepts or rejects in a JSON log. For a business owner, that means your AI processes become auditable. If your agent suddenly starts handling something incorrectly, you can trace exactly which instruction was added or removed. You can also hand that file to an employee or consultant for review. In a regulatory climate where you might need to explain how you processed customer data, that’s valuable.

Here’s the key cross-harness data from the research, simplified:

Skill Trained In	Deployed In	Baseline (No Skill)	Direct (Trained In Target)	Transferred
Codex	Claude Code	22.1	80.4	81.8
Claude Code	Codex	27.5	85.0	71.1
Claude Code	Codex (LiveMath)	35.2	78.4	48.0
Codex	Claude Code (LiveMath)	40.8	56.5	42.4

The first row is the standout: the transferred skill exceeded the score Claude Code achieved by training directly on its own. That tells us the skill isn’t just a collection of tool-specific commands. It’s a reusable procedure. The LiveMath rows at the bottom, however, show that math-heavy reasoning skills transfer weakly—only 10% of the in-domain gain survived. So portability isn’t universal. It depends on how procedural vs. reasoning-heavy the task is.

Practical Takeaways

So what should you do today, as an SME owner? Don’t wait for SkillOpt to be released. Start thinking about your AI workflows differently—as portable assets, not platform features.

• Write your AI procedures as plain language. The most portable skill format is a simple text document with clear, step-by-step instructions. Start maintaining a “skills folder” for your business where each procedure is documented in text.
• Separate “how” from “what.” If you’re instructing an AI to handle your inventory reconciliation, make the skill about the process (check stock levels, match receipts, flag discrepancies) rather than about the specific tool (click this button, run this command).
• Test portability when you can. If you already use tools like ChatGPT and Claude, try taking one of your existing prompt systems and running it in the other tool with minimal changes. Note what breaks and what survives. That tells you how dependent your current setup is on the platform.
• Inspect before you deploy. When you set up an AI workflow, ask for a traceable log of what the agent did and why. This habit will pay off as more AI systems become transparent about their internal steps.

“Portability plus inspectability is a different operational posture than shipping fine-tuned weights.” — This is the core lesson for business owners: your AI knowledge can be a written, reviewable, and movable document instead of a locked-in, invisible process.

The Bigger Picture

Long-term, this trend points toward a future where AI agents are more like interchangeable employees than proprietary tools. Just as you’d hire a new staff member and give them your company’s operations manual, you might soon give an AI agent a skill document that works regardless of which “brain” it runs on. That’s a good thing for competition, and it’s a good thing for you—it means lower switching costs, more choice, and more direct control over how your business processes are defined.

The research is still early. It covers only one model family (GPT) and a handful of benchmarks. Cross-family transfer—like moving from GPT to Qwen—isn’t tested yet. But the direction is clear: AI knowledge is becoming portable. For a Malaysian SME, that’s one more reason to start documenting your AI workflows today. Because when portability arrives, the businesses that already have clean, well-written skill documents will be the ones who can move freely.

Ready to Streamline Your Operations?

Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →