Everyone has an opinion about AI tools. Most of those opinions come from people selling them.

Here is the uncomfortable reality: the AI tool landscape has become a swamp of overlapping features, competing subscriptions, and promises that rarely survive contact with actual work. Building a useful tool stack requires ignoring most of what you read about building a useful tool stack.

The Collector’s Trap

You start with ChatGPT. Then someone mentions Claude writes better. You sign up. Then you hear Perplexity is better for research, so you add that too. Midjourney for images, obviously. Jasper for marketing copy because it has templates. Notion AI because you already use Notion. Gemini because it integrates with your Google stuff.

Six months later, you have seven subscriptions and the vague sense that you are not actually more productive.

This is normal. A study from the Cerbos engineering blog found that “teams with high AI adoption interacted with 9% more tasks and 47% more pull requests per day” but were not necessarily completing more meaningful work. The activity feels productive without translating into results that matter.

As one Hacker News user put it bluntly: “For me it’s just a glorified stack overflow.”

That is not dismissive. It is honest. And honest assessments of AI tools are rare because honesty does not sell subscriptions.

The Single Tool vs Stack Question

There is a persistent debate in online communities about whether you should master one AI tool deeply or spread across several specialized ones. The answer depends on something people rarely mention: the friction cost of context switching.

When you bounce between ChatGPT, Claude, and Perplexity within the same task, you lose time. Not just the seconds of switching tabs, but the mental overhead of remembering which tool you have open, what context you have already provided, and what each tool does slightly differently from the others.

A Hacker News user named dexterlagan described the breakthrough moment clearly: “I’ve been writing detailed specs to direct LLMs, and that’s what changed everything for me.” Notice he did not say switching to a better model changed everything. He said changing how he worked with the tools changed everything.

This points to an uncomfortable truth. How you use a tool matters more than which tool you use, for most people, most of the time, across most tasks.

The Category Breakdown Nobody Asked For

Writing tools split into several overlapping buckets: general purpose assistants like ChatGPT and Claude, specialized copy tools like Jasper and Copy.ai, and editing tools like Grammarly that now have AI features bolted on.

Image generation similarly fragments. DALL-E lives inside ChatGPT. Midjourney produces distinctive aesthetic results but requires using Discord, which some people hate with the intensity of a thousand suns. Stable Diffusion runs locally if you have the hardware and patience. Leonardo.ai offers a web interface with more control than DALL-E but less distinctive output than Midjourney.

Coding assistants form their own category. GitHub Copilot, Cursor, Cody, Tabnine, and a dozen others all promise to write code faster while introducing bugs in ways you have never seen before. The Cerbos research found that “AI-generated code introduced 322% more privilege escalation paths and 153% more design flaws” compared to human-written code. Speed and quality pull in opposite directions.

Research tools present yet another category. Perplexity searches the web and cites sources. ChatGPT added web search. Claude has a research mode. Google’s Gemini integrates with your actual data. Each handles the same fundamental task with different strengths and weaknesses and limitations and blind spots.

The categories blur. ChatGPT writes copy, generates images, analyzes data, and browses the web. Claude does most of those things. Trying to maintain strict category boundaries feels like organizing fog.

How Tools Work Together (Or Don’t)

The marketing around AI tool stacks suggests seamless integration. Reality is messier.

ChatGPT cannot see what you made in Midjourney. Claude does not know what Perplexity found. Your CRM has AI features that ignore the context from your email tool, which has different AI features ignoring context from your CRM. Each tool operates in isolation, forcing you to manually shuttle information between them.

Some integrations exist. Zapier connects things. Make.com connects things. n8n connects things if you enjoy YAML. But the integrations typically handle simple triggers and actions: when this happens, do that. The nuanced back-and-forth that makes AI useful resists easy automation.

The result is that your “stack” often becomes a set of disconnected tools you use separately rather than a coherent system that amplifies your capabilities. You have multiple AI assistants who have never met each other and never will.

This isolation carries a hidden cost. You provide context to one tool, get a result, then manually transfer that context to another tool and provide additional context on top. The duplication adds up across a week, a month, a year of working this way.

The Integration Landscape

Native integrations offer the smoothest experience. Notion AI within Notion, Grammarly within Google Docs, Copilot within VS Code. The tool lives where you already work, using context it already has, without requiring you to copy and paste anything.

The tradeoff is capability. Native integrations typically lag behind standalone tools in features and model quality. Notion AI is convenient but limited. Grammarly’s AI suggestions pale compared to Claude’s writing abilities. Convenience and power trade off against each other.

Third-party connectors fill some gaps. Zapier has 7,000+ app integrations. Make.com offers similar breadth with different pricing. Both let you create automation workflows that pass data between tools automatically.

But there is a meaningful difference between “automatically triggers when” and “intelligently responds to.” Automation handles the mechanical parts well. The judgment parts still require human intervention or manual prompting.

When Less Actually Is More

Counterintuitive finding: some of the most productive knowledge workers use fewer AI tools, not more.

The logic works like this. Every tool requires learning its quirks, understanding its limitations, discovering its hidden features, and building habits around its interface. That investment compounds over time. Deep familiarity with one tool often produces better results than shallow familiarity with five.

A Hacker News commenter named joshstrange captured this: “Copilot is the sweet spot…It saves me significant time when coding.” Not multiple coding assistants. One tool, used consistently, producing consistent value.

The minimalist approach also reduces decision fatigue. When you have one writing tool, you open that tool and write. When you have three writing tools, you first decide which tool to use, then worry whether you chose wrong, then maybe switch halfway through. The choosing itself consumes energy.

There is also a cost floor. More tools mean more subscriptions, which means more money leaving your account monthly. At some point, the marginal value of an additional tool fails to justify its marginal cost. Many people blow past that point without noticing.

The Honest Productivity Question

Does AI actually make you more productive? The research gives a conflicted answer.

Developers using AI were on average 19% slower in one study, yet they were convinced they had been faster. The feeling of productivity diverged from the reality of productivity. AI feels fast because it gives instant feedback. Type a prompt, get a response. The reward loop activates regardless of whether the response actually helps.

Only 16.3% of developers in another survey said AI made them more productive “to a great extent.” The largest group, 41.4%, said it had little or no effect. These numbers do not match the marketing.

The problem may be a version of the 70% problem that practitioners discuss. AI can get you 70% of the way, but the last 30% is the hard part. And that last 30% often takes as long as doing the whole thing would have taken, negating the time savings from the first 70%.

For writing specifically, AI excels at generating first drafts that require substantial editing. Is draft generation plus heavy editing faster than writing from scratch? Sometimes yes, sometimes no. Depends on the person, the task, the quality bar.

Building Your Stack (If You Insist)

Start with one general-purpose assistant. ChatGPT or Claude. Both work. Pick one and use it for a month before adding anything else.

Track where it fails you. Not where it is theoretically limited, but where it actually blocks your specific work. Those failure points identify where a second tool might add value.

Add a second tool only when you have a clear use case the first tool cannot handle. Maybe image generation. Maybe code completion. Maybe research with citations. One specific gap, one specific tool to fill it.

Resist the third tool. Seriously. Two tools that you use constantly beat five tools that you kind of use sometimes. The switching costs accumulate faster than the benefits.

If you must add more, review your stack quarterly. Audit actual usage, not theoretical utility. Cancel subscriptions for tools you have not opened in thirty days. The recurring charges continue whether you use the tool or not.

The Stack That Probably Works

For most knowledge workers, this works:

One AI assistant. ChatGPT Plus or Claude Pro. Handles writing, analysis, brainstorming, and basic image generation. Covers 70% of AI use cases.

One image tool (if needed). Midjourney if quality matters, DALL-E if convenience matters. Or skip entirely if images are not central to your work.

One coding assistant (if you code). Copilot or Cursor. Integrated into your editor. Available without context switching.

That is it. Three tools maximum. Total cost under $100 per month. More capability than you will probably use.

The temptation to add more will persist. You will read about some tool that sounds perfect for some workflow. You will sign up for the free trial. You will use it twice and forget it exists. This is the cycle.

The Uncomfortable Conclusion

AI tools are getting better quickly. The stack you build today may be obsolete in a year. The specific recommendations matter less than the underlying principle: do not collect tools like trading cards hoping they will someday be valuable.

Find the smallest set of tools that handle your actual work. Use them consistently enough to develop real skill. Ignore the rest.

As one commenter observed about tool complexity: “Your workflow is not sacred…If your workflow isn’t changing…you have grown stagnant.” The tools will change. Your needs will change. Building an elaborate system around today’s tools may just mean more work rebuilding when tomorrow’s tools arrive.

The best AI tool stack might be the one you think about the least.