Your AI tools just got more practical — here’s why
As a Malaysian SME owner, you’ve probably spent hours crafting the perfect prompt for ChatGPT, Claude, or another AI assistant. Maybe you’ve tuned it to generate sales reports, reply to customer inquiries, or calculate your monthly payroll. Now imagine you could move that same instruction set to a completely different AI tool — with no retraining and no loss in performance. That is exactly what Microsoft’s new SkillOpt research demonstrates. And it matters more to you than any flashy AI launch, because it speaks directly to the biggest practical headache in business automation: vendor lock-in.
What Happened
Researchers from Microsoft, Shanghai Jiao Tong University, Tongji University, and Fudan University developed SkillOpt, an optimizer that trains a single natural-language skill document while the target model stays frozen. The exported artifact is one file called best_skill.md. This approach is fundamentally different from fine-tuning, which modifies the model’s internal weights. Here, the model itself never changes — only a text instruction set is optimized.
The team then tested whether these skill documents transfer. In cross-model experiments within the GPT-5.4 family, a skill trained on the full model and deployed on GPT-5.4-mini kept 82% of the in-domain gain on SpreadsheetBench. More striking were the cross-harness tests. A skill optimized inside OpenAI’s Codex lifted Claude Code’s SpreadsheetBench score from 22.1 to 81.8 — slightly above the 80.4 Claude Code achieved when trained on its own skill from scratch. These two harnesses expose completely different tool and file APIs, yet the skill still transferred.
“The two harnesses expose different tool and file APIs and different command surfaces. A skill that survives that shift is not encoding command recipes.”
The research paper explains that SpreadsheetBench’s portability comes from workbook-level procedures: structure-first inspection, formula-aware verification, and static-value materialization. Those procedures hold regardless of which CLI runs the Python code. But not everything transfers equally — LiveMath, a reasoning-heavy benchmark, retained only 10% of the in-domain gain when moving from Codex to Claude Code. Procedural skills are portable; reasoning-heavy skills are not.
Why This Matters for Malaysian SMEs
You don’t need a research lab to benefit from this. Think about a typical local business scenario: you use an AI assistant to process supplier quotes in Excel, generate purchase orders, or standardize your monthly inventory count. You’ve likely discovered that a prompt that works beautifully in ChatGPT behaves differently when you paste it into Claude or a local automation tool built on an open-source model. SkillOpt’s finding means your carefully refined instructions can live independently of the platform. You could build a skill on the tool your staff finds easiest to use, verify it, then deploy it where the actual work happens — even if that means switching AI providers next year.
There is also an audit advantage that you hear too little about. The deployed artifact is a text file a domain practitioner can read in minutes. Every change to it is traceable — each edit step records an edit_apply_report.json with per-edit accept and skip status. For your business, that means you don’t need to trust a black box. If your bookkeeping assistant uses a skill to prepare GST data, you can open the file and see exactly what steps it takes before outputting anything. That level of transparency matters for internal controls and for answering a customer or auditor who asks, “Why did your system make that decision?”
The Bigger Picture
What SkillOpt shows is that training cost is paid once, offline, and measured. The optimizer model runs only during training and adds zero inference-time calls at deployment. For your SME, that suggests a practical strategy: standardize your core business procedures into portable skill files, verify them once, then reuse them across every AI tool you adopt. This is a different operational posture than shipping fine-tuned weights, which are expensive to create and impossible to inspect.
Still, the evidence has limits. All transfer results stay inside one GPT family, and only three transfer axes were tested. As the article notes, portability is demonstrated, not yet generalized. The lesson for you is to start with procedural tasks — spreadsheet processing, document formatting, data extraction — rather than complex mathematical reasoning. Those procedural skills carry their value across environments, while reasoning skills remain tied to their training environment.
Key transfer numbers at a glance
| Source → Target | Baseline | Transferred | Gain |
|---|---|---|---|
| SpreadsheetBench: Codex → Claude Code | 22.1 | 81.8 | +59.7 |
| SpreadsheetBench: Claude Code → Codex | 27.5 | 71.1 | +43.6 |
Source: MarktechPost
- Start with procedural tasks: Spreadsheet workflows, document generation, and data verification transfer well across models and harnesses.
- Keep your skills in plain text: A skill file of a few hundred to a few thousand tokens can be read, edited, and audited by your team.
- Test before you commit: Transfer is not universal — reserve math and reasoning-heavy automations for the tool you originally trained them on.
- Build once, reuse many times: One optimized skill can serve ChatGPT, Claude, and future tools you haven’t chosen yet.
The practical takeaway is straightforward. You don’t need to be a machine learning expert to build durable business automation. You need clear, procedural instructions that you control. With research like SkillOpt proving those instructions travel well, your next AI project can be built with the freedom to switch tools whenever you want. That is the kind of flexibility a growing Malaysian SME should plan for.
Ready to Streamline Your Operations?
Technology moves fast. Your operations should keep up. AutoRunBiz builds AI systems that run your daily workflows — from WhatsApp order capture to accounting. Book a free 15-min ops audit →