The Skill Shift: Why GPTs Died and Mercury Flux Was Built for What's Next

TL;DR: Google is killing Gems. OpenAI just killed personal GPTs for individual accounts. The era of "wrap a prompt in a chat window, add an avatar, share a link" is officially over. What's replacing it? Skills—lightweight, self-maintaining, team-sharable workflows that agents execute end-to-end. Mercury Flux was built for this exact shift: route the right model to the right task, with local memory that persists across sessions. The model leaderboard changes weekly. Your workflow shouldn't have to.

James here, CEO of Mercury Technology Solutions. Wanchai, Hong Kong — August 22, 2026

Two announcements landed this week that should have every AI "product builder" sweating.

Google: Gems is shutting down October 20. Export your stuff or lose it.

OpenAI: Personal accounts can no longer create or publish GPTs. Existing ones work—for now. But the "build a GPT, share the link, call it a product" pipeline is dead.

If you built your AI "business" on GPTs or Gems, you're holding deprecated inventory. If you taught courses on how to build them, your curriculum just aged out. If you outsourced your knowledge work to a custom GPT that you manually updated and copy-pasted from... well, we need to talk.

Why the Giants Are Killing Their Own Toys

This isn't capricious. It's structural.

When OpenAI renamed the basic chat interface to "Classic" and started pushing ChatGPT Work and Codex, they signaled the shift: from conversational AI to agentic AI.

The old model: You ask. AI answers. You execute.

The new model: You delegate. AI executes. You verify.

GPTs and Gems were training wheels—useful for proving that non-technical users could "build" something with AI. But they were never serious infrastructure. They couldn't self-maintain. They couldn't integrate with your tools. They couldn't hand off to team members. They were, essentially, prompts in a prettier box.

Skills are different. Skills are workflows with judgment boundaries—input specifications, output standards, steps that can't be skipped, checkpoints that require human approval. They're lighter than GPTs, more maintainable, and designed for teams rather than solo users.

The giants aren't abandoning the space. They're upgrading the abstraction layer. And if your business is still sitting at the old layer, you're about to experience abstraction collapse.

The 2026 Equation: Prompt wrappers = rented workflows. Skills = owned infrastructure. The difference is policy risk.

The 2026 Equation: Prompt wrappers = rented workflows. Skills = owned infrastructure. The difference is policy risk.

The Model Leaderboard Is a Red Herring

Here's something else that happened this week: another model ranking shuffle.

Fable > Grok 4.6 > Kimi 3 > Opus 4.8 > Grok 4.5 > Codex 5.5 > Qwen 3.8 > Kimi 2.7 > Opus 4.6 > GLM 5.2 > minmax

Catch your breath. It'll change next month.

The obsession with "which model is best" misses the point. Different models excel at different tasks. Grok is great at reasoning. Kimi 3 has massive context. Opus handles nuance. Codex is built for software. Qwen is efficient for Asian languages. The "best" model is the one that solves your specific problem at the lowest cost with acceptable quality.

This is why we built Mercury Flux.

Flux doesn't bet on one model. Flux routes. It maintains local memory of your tasks, your preferences, your outputs—and sends each job to the model that's actually best for it. Complex reasoning? Grok. Long-document analysis? Kimi. Code generation? Codex. Quick classification? Qwen. All without you manually switching tabs or copying context between windows.

The model leaderboard is entertainment for AI Twitter. Model routing is infrastructure for people who actually work.

From GPTs to Skills: What Changed

Let me make this concrete with a real example.

Before (GPTs Era)

I had two GPTs: one for writing, one for "de-AI-ing" content.

Every few weeks, I had to manually update their knowledge bases with my latest articles, current projects, new writing techniques I'd learned. The GPTs would generate drafts, then I'd copy-paste into email or WordPress, reformat, add links, check tone.

The AI gave me first drafts. All the maintenance, integration, and finishing work was on me.

After (Skills Era)

I wrote my methodology as a Skill. Not a prompt—a workflow.

My AI agent now:

1. Self-updates its knowledge base from my social posts, blog, newsletter, and work logs
2. Generates drafts using my current context and voice
3. Self-edits for tone ("de-AI-ing" is built into the workflow)
4. Delivers to Gmail drafts or WordPress editor, within authorized boundaries
5. Presents finished work for my approval

I do what humans should do: review, tweak, publish. The agent does what it should do: everything else.

Same "writing assistant." Completely different job description. The GPT was a tool I operated. The Skill is an employee I manage.

The Local Memory Imperative

Here's the part most people miss.

When you build on someone else's platform—GPTs, Gems, Claude Projects, whatever—you don't own your workflow. You rent it. And the landlord can change the terms, raise the rent, or demolish the building.

Google just proved that. OpenAI just proved that.

This is why I've been telling everyone: put your knowledge, your skills, your reference materials in local storage that you control. Not because cloud is bad, but because policy risk is real.

With Mercury Flux, your local memory persists across models. If OpenAI changes pricing tomorrow, you switch the engine to Kimi. If Kimi goes down, you route to Grok. Your workflow—your Skills, your context, your judgment boundaries—stays yours.

The AI companies may not survive 10 years. Your skills should.

What This Means for Your Business

If you're an AI trainer, course creator, or consultant who taught people to build GPTs: pivot. Now. The skills you taught aren't worthless, but the container is dead. Teach Skills. Teach workflow design. Teach how to decompose a job into steps that an agent can execute with appropriate human checkpoints.

If you're a business that outsourced work to custom GPTs: migrate. The GPTs will keep working for a while, but you're on deprecation notice. Build Skills that integrate with your actual tools—your CRM, your CMS, your email, your analytics.

If you're a knowledge worker who thought AI would be a chat interface forever: wake up. The chat was the tutorial level. Agentic execution is the main game.

The Action

Stop building prompts. Start building workflows with judgment.

1. Inventory your GPTs/Gems. What do they actually do? Can you describe the workflow in steps?
2. Define boundaries. What requires human approval? What can the agent decide autonomously?
3. Choose routing over loyalty. Don't marry a model. Marry a workflow that can use any model.
4. Local-first your knowledge. Your IP should survive any platform's policy change.
5. Test Mercury Flux. If you're running multiple models for different tasks, you're already doing manual what Flux does automatically.

The giants just declared that the prompt-wrapper era is over. They're not wrong. They're just late to admitting what practitioners already knew: real work requires real workflows, not chat windows with avatars.

Build accordingly.

Related Mercury surfaces:

• Mercury Flux — Route every task to the cheapest capable model. Local memory. No lock-in.
• Mercury Core — The OS where humans and agents share memory. Skills live here.
• Talk to Your ERP — Operations through conversation, on Core.
• AI consulting — Architecture, not a tool bake-off.

Mercury Technology Solutions: Accelerate Digitality.