AI Is Just Okay, and the Stuff It Writes Is Unusable

"AI is just okay. The stuff it writes is unusable."

"ChatGPT? I tried it. Not useful at all. I still have to do it myself."

I've said both of those things out loud. Not this year—late 2023, when ChatGPT first blew up in China. I used it to draft a client email, revised it three times, and it still had that plastic, off-the-shelf feel. I genuinely thought the whole thing was overhyped at the time.

January 23, 2026. Friday. I sat back down, thought it through, and decided to lay this out in the open: you think AI doesn't work—what if the problem is that your way of using it is still stuck three years in the past?

A Few Data Points I Keep Coming Back To

An analysis on the Alibaba Cloud Developer Community put it this way:

"If 2023 was the dazzling debut of large language models, then 2026 will be defined as the first year of Agents."

Gartner's forecast is even more direct:

"By the end of 2026, 40% of enterprise applications will have AI Agents embedded—up from less than 5% in 2025."

Data from AI market research firms points in the same direction:

"The Agentic AI market will grow from $7.8 billion to $52 billion by 2030."

The numbers are right there, and I still hear people telling me "AI doesn't work." I've turned it over in my mind a lot, and the answer is actually pretty simple: they're still talking to 2026's AI the way they talked to 2023's.

Four Years, Four Shifts in How I Use It

I went back and mapped out these four years on a timeline. It looks roughly like this:

2023: I Ask, You Answer

ChatGPT shook the world. A lot of people discovered for the first time that AI could actually hold a conversation like a human.

My usage back then was: throw a question, wait for an answer, close the tab. "Write me an email." "Explain quantum mechanics." "Tell me a joke." That was the whole menu.

The limitations were glaring. Single-turn conversations with no real context carryover. Knowledge with a hard cutoff date. No access to external tools. And it would confidently spout nonsense on a regular basis.

DigitalOcean's analysis nailed it:

"AI chatbots are designed for simple, single-intent interactions. When a query requires multi-step reasoning or involves layered instructions, the chatbot typically falls back to a generic answer or human assistance."

To put it bluntly, most people I know are still stuck at this level today.

2024: You Orchestrate, I Execute

Conversation alone couldn't carry the workload, and everyone figured that out fast. Two things emerged:

**Workflow**: Break complex tasks into steps and chain multiple AI calls into a single pipeline. LangChain was the hottest framework that year. I tinkered with it too—honestly, getting a chain to run was doable, but changing a requirement meant half a day of rework.

**RAG (Retrieval-Augmented Generation)**: AI no longer answered purely from "memory." It could query external knowledge bases in real time. Hallucinations dropped significantly. I tested this myself, and the improvement was immediate.

Typical use cases: customer service bots connected to FAQ databases, document Q&A systems, multi-step data processing pipelines.

Where did it fall short? The pipeline had to be designed by specialists. Once the flow was locked in, it was rigid. Hit an unexpected situation, and it couldn't adapt.

A technical analysis on Medium explained it clearly:

"Traditional RAG passively retrieves and generates responses. Agentic RAG adds goal-driven reasoning on top—turning the AI into an Agent that can plan steps, retrieve multiple pieces of information, use tools, and reflect on its progress."

2025: MCP, Connecting Everything

In November 2024, Anthropic released MCP (Model Context Protocol).

My mental model at the time: MCP is the USB-C port for AI. Just as USB-C standardized how you connect electronic devices, MCP standardized how AI connects to external systems.

What did that mean in practice? Before, wiring AI to three tools meant writing three separate integration scripts. Now, as long as a tool supports MCP, AI can use it out of the box. I only truly felt the difference after hitting a wall myself—I'd manually written a GitHub API integration script, and the moment MCP came out, that code was obsolete. Plug in an MCP server, done in five minutes.

The scale, per MCP's official data:

"One year in, MCP has become the universal standard for connecting AI Agents to enterprise tools—over 97 million SDK downloads per month, with support from Anthropic, OpenAI, Google, and Microsoft."

The ecosystem exploded fast: over 5,800 MCP servers, over 300 MCP clients. From GitHub to Notion, from Slack to databases—everything was connecting.

2026: Agent + Skill, Autonomous Execution

Two concepts are carrying 2026.

**AI Agent**. An analysis in the cross-border e-commerce space drew the distinction sharply:

"The key difference: an AI Agent has something a Chatbot doesn't—Agency. That means it can make decisions and take real action, not just chat."

What does an Agent actually do? Plan (break a big goal into steps), Execute (call tools step by step), Reflect (evaluate results and self-correct), Loop (keep going until the task is done).

**Agent Skills**. At the end of 2025, Anthropic released the Agent Skills specification, and OpenAI and others followed suit. The official definition:

"Agent Skills are organized folders of instructions, scripts, and resources that an Agent can dynamically discover and load to perform specific tasks more effectively."

There's a key design principle behind it called Progressive Disclosure: Layer 1 loads only the skill's name and description; Layer 2 loads the full documentation when needed; Layer 3 loads the relevant scripts and resources when needed. This design solves the problem of tool descriptions blowing up the context window. Now AI loads on demand, so in theory there's no upper limit on the number of skills.

By 2026, my workflow looks like this: I set the goal. The Agent plans the steps, calls the skills, executes, and adjusts on its own. My job is to show up only at the critical decision points.

One Table to Make the Difference Clear

Strip it down to the core, and the shift is one sentence: from "you tell AI how to do it" to "you tell AI what to do."

A passage from an Agentic AI trend analysis stuck with me:

"In 2024, Copilot was the co-pilot—it offered suggestions, but you were still on the wheel. You prompted, reviewed, edited, pasted. In 2026, autonomous AI workers operate 'human on the loop' or even 'human out of the loop.'"

The Pitfalls I've Hit—And Why You Think AI Doesn't Work

Let's get real. From what I've observed, plus my own mistakes, the reasons basically boil down to four.

**Still using it like a Q&A machine.** I made this mistake early on. I'd throw out "Write me an article about AI" and get back something that read like it could've been written by anyone—correct but empty. Then I switched to this:

"I'm a tech blogger. My target audience is people who are interested in AI but aren't deeply technical. Write me an article about how AI usage has evolved, from 2023 chatbots to 2026 Agents. Use a storytelling approach so readers understand why they need to update how they use AI. Include specific examples, data-backed arguments, and actionable advice. For style, refer to this example: [attach a previous article]"

What's the difference? The first is a "question"—vague in, vague out. The second is a "task brief"—AI knows who you are, who it's writing for, what the goal is, and what the style should be. In my experience, the second approach produces a first draft that saves me about half my revision time.

**Not giving AI enough "tools."** If you're only using the ChatGPT web interface: it can't touch your files, can't search for the latest info, can't run your code, can't connect to your tools. Switch to Claude Code + MCP, and it can read and write your codebase, execute shell commands, connect to databases, call various APIs, and chain complex tasks end to end. Same model, different "loadout," and the results can be worlds apart. I'm not exaggerating—I ran the same article through the web version and through Claude Code and compared the output.

**Not giving AI any "skills."** Here's an example. I asked Claude to write code that follows our company's coding standards. Without a skill installed, Claude had no idea what our standards were and just went with generic best practices—out came a pile of code with the wrong style. After I packaged the standards into a Skill, Claude loaded it on its own, followed it, and the code came out basically format-clean with no fixes needed from me.

Anthropic's official description puts it well:

"Skills allow users to package domain expertise into composable resources, transforming a general-purpose Agent into a specialized one."

**Expectations set in the wrong place.** My biggest early mistake was expecting AI to deliver a perfect answer in one shot. When it didn't, I wrote it off as "not good enough." The reasonable expectation: AI is a collaborative partner. Answers need iteration and adjustment. A Claude Code analysis on Medium put it nicely:

"Interestingly, Claude Code isn't the most popular coding Agent because of flash. It's because of reliability, thoughtfulness, and scalability."

AI is a tool, not magic. Tools only deliver when used right.

My Current Setup and Workflow

Here are four things I actually use. Feel free to copy them.

**Upgrade from "Q&A" to "task brief."** Don't ask: "How do I read a CSV file in Python?" Say this instead: "I have a file called sales_data.csv with three columns: date, product name, and sales amount. Write a Python script that reads the file, aggregates sales by month, and generates a bar chart saved as monthly_sales.png." The difference: in the first case, you ask about a knowledge point and get a knowledge point back. In the second, you hand over a task and AI completes the whole task. When I'm writing scripts, it's almost always the second approach, and the time saved is real.

**Give AI a full "toolkit."** Minimum setup: use Claude.ai or ChatGPT Plus, turn on web search and code execution. Step up: get Claude Code so AI can work with your file system; set up MCP servers to connect your everyday tools to AI. My own setup (for reference): primary workbench is Claude Code; MCP connected to GitHub, databases, Notion, and others; common workflows packaged as custom Skills. Once this stack is running, my daily coding and writing workflow is completely different.

**Build your own "Skills library."** What are Skills, in one sentence: package your domain knowledge, workflows, and best practices into a format AI can read and use. How to build one? Create a folder named after the skill; write a SKILL.md file explaining what the skill does; drop in any related docs, scripts, and templates; put the folder in Claude Code's skills directory. The official docs mention:

"Skills are installed from the anthropics/skills marketplace via plugins. When relevant, Claude loads them automatically."

I've got a dozen or so Skills now—everything from code review standards to article writing templates to data report formats. All built by me.

**Learn to "direct," stop "operating."** Traditional way: you write code → you test → you fix bugs → you test again. Agent way: you describe the goal → Agent writes code → Agent tests → Agent fixes bugs → you review the result. Your role shifts from "doer" to "director." This has nothing to do with being lazy—it's a fundamental shift in efficiency. I've mentored junior devs. Before, I walked them through everything hand in hand. Now I describe the requirement clearly, let the Agent produce a first pass, I review it, and then the junior dev looks at it. The efficiency is on a completely different level.

Two Real Scenarios I've Run

Writing This Article

The traditional way—I estimated the time: research 2 hours, outline 1 hour, first draft 3 hours, revision and polish 2 hours, images 1 hour. Total: 9 hours.

The Agent way: I give the topic and requirements. Claude Code searches for the latest material. Claude Code builds the outline. Claude Code writes the first draft. I review and edit. Claude Code generates the images. Total: 2 hours.

Efficiency gain: 4–5×. I didn't just guess that number—I timed it.

A Data Analysis Report

The traditional way, end to end: write SQL to pull the data, export to Excel, clean it in Python, chart it with matplotlib, write a Word report, then convert to PPT. Last time I did this the old way, it took a full afternoon.

The Agent way: I just dropped one line:

"Analyze last month's user behavior data. Identify the three conversion stages with the lowest rates. Provide optimization recommendations. Generate a PPT report."

Claude Code took it from there: connected to the database, ran the queries, cleaned and analyzed the data, generated visualizations, wrote the analysis report, and output the PPT file.

All I had to do: review the results and give feedback. From a full afternoon down to forty minutes.

A Few Final Thoughts

Three years ago, I was the person slamming the table and saying "AI doesn't work." ChatGPT articles screamed "AI-generated." AI-drawn images were crooked. AI-written code couldn't even run.

Then I finally admitted it: AI wasn't the problem. I didn't know how to use it.

Same sharp knife. One person can only slice potatoes. Another can carve intricate patterns.

2026's AI has a completely different build: not just a chatty Chatbot, but an Agent that plans and executes autonomously, an MCP that connects everything, and Skills that scale without limit. Still using the 2023 posture? Of course it feels underwhelming.

Key Takeaways

• AI usage has gone through four stages: 2023 Chatbot → 2024 Workflow + RAG → 2025 MCP → 2026 Agent + Skill
• The core shift: from "you tell AI how to do it" to "you tell AI what to do"
• Why most people think AI doesn't work: still in Q&A mode, no tools given to AI, no skills given to AI, wrong expectations
• The right approach: task briefs instead of Q&A, set up tools and MCP, build a Skills library, learn to direct instead of operate
• Key 2026 data points: 40% of enterprise apps will embed AI Agents; market size growing from $7.8 billion to $52 billion

January 23, 2026. Friday. General-purpose AGI tools have arrived. The question isn't whether AI works—it's whether you're ready to update your "posture."

References

Official Resources

• Model Context Protocol - Anthropic - Official MCP release
• Agent Skills - Anthropic - Skills technical deep dive
• Agent Skills Overview - Claude Docs - Official Skills documentation

Trend Analysis

• 7 Agentic AI Trends to Watch in 2026 - Machine Learning Mastery
• The Era of Agentic AI - Downgraf
• Agentic AI Trends 2025 - Svitla Systems

Technical Explainers

• RAG vs Agentic RAG vs MCP - Bitcot
• AI Agent vs Chatbot - DigitalOcean
• 2026: The First Year of Agents - Alibaba Cloud Developer Community

This article is based on publicly available information as of January 23, 2026. The evolution of AI usage is still ongoing. Today's "right posture" may be surpassed tomorrow—and that's exactly what makes this era so exciting.

Follow me for more insights, cognition, and practical approaches to AI Agents. In an era where AI is evolving this fast, the only way to stay ahead is to keep learning and keep updating your mental models.