Tom Hundley
🤖 Ghostwritten by Claude Opus 4.5 · Curated by Tom Hundley
This article was written by Claude Opus 4.5 and curated for publication by Tom Hundley.
I have a confession: I used to be one of those people who tried every new AI tool the moment it launched. Browser tabs multiplied. Subscriptions stacked up. My productivity actually decreased because I spent more time evaluating tools than using them.
That phase is over.
What you're about to read isn't a vendor comparison or a "best of" list. It's simply what I use. Every day. The tools that survived my ruthless pruning, the ones I reach for without thinking, and—just as importantly—where each one falls short.
Here's the uncomfortable truth nobody in the AI space wants to admit: we're not there yet. These tools are genuinely transformative, and I wouldn't want to work without them. But they're also frustrating, inconsistent, and sometimes hilariously wrong. Both things are true.
This series documents my digital assistant stack as it exists today. I'll share what works, what doesn't, and how these tools actually fit together in practice. Consider it a practitioner's field notes rather than a polished recommendation.
Before diving into the specifics, let me explain how I think about building a tool stack.
The temptation with AI tools is to find "the one"—a single platform that handles everything. That's the wrong mental model. General-purpose tools optimize for breadth, not depth. They're jacks of all trades, masters of none.
Instead, I think about tools in layers. Each layer handles a specific type of work:
No single tool dominates every layer. Claude excels at nuanced reasoning and now has web search capabilities. ChatGPT has broader integrations and a more mature ecosystem. Perplexity finds information fast with inline citations. The stack works because each tool has a specific job.
This is where I spend most of my AI interaction time—thinking through problems, getting feedback on ideas, and working through complex decisions.
Claude has become my default for any task requiring nuanced thinking. When I need to analyze a business problem, draft a difficult email, or work through a complex technical architecture, I open Claude first.
What earned Claude this position isn't just intelligence—it's the quality of engagement. Claude pushes back when I'm wrong, asks clarifying questions, and admits uncertainty. I've found this leads to better outcomes than tools that simply try to give me what I want.
Claude Desktop added some important capabilities recently. The ability to create and edit documents directly—including Excel spreadsheets and PowerPoint presentations—means I can stay in context rather than copying outputs into other applications. The Desktop Extensions feature, built on Model Context Protocol (MCP), lets Claude connect to external tools. I have it connected to my local files and Notion, which means Claude can actually reference my existing work rather than starting from scratch every conversation.
The web search capability, which rolled out globally in May 2025, addresses what used to be Claude's biggest limitation. When I need current information—recent news, live documentation, today's stats—Claude can now search the web directly. This has meaningfully changed how I use it for research tasks.
The "Projects" feature helps too. I maintain ongoing projects for different work streams, and Claude remembers the context between sessions. This sounds small, but it eliminates the tedious re-explanation I used to do at the start of every conversation.
Where Claude falls short: While web search has addressed the knowledge cutoff issue, the results can sometimes be less comprehensive than dedicated research tools like Perplexity. For deep-dive research across many sources, I still reach for specialized tools.
ChatGPT still earns its place in my stack, though its role has shifted over time.
I reach for ChatGPT when I need integrations that Claude doesn't have. With direct connections to Gmail, Google Calendar, and Google Contacts, ChatGPT can do things like summarize my calendar for the week or draft responses that reference specific email threads.
ChatGPT also shines for multimodal work. When I need to analyze an image, discuss a diagram, or get feedback on a visual design, ChatGPT's vision capabilities are mature and reliable. The voice mode is surprisingly useful for hands-free conversations—I'll sometimes use it while walking or driving.
The shopping assistant is unexpectedly helpful. No ads, side-by-side comparisons, and actual links to products. I'm skeptical of AI-as-shopper in general, but for research-heavy purchases it saves real time.
Where ChatGPT falls short: For deep, nuanced thinking, I find the outputs more generic than Claude's. ChatGPT tends to give me what it thinks I want rather than pushing back or surfacing complications. It's also more prone to confident-sounding errors.
Perplexity occupies a specific niche: when I need to learn something fast with sources I can verify.
Unlike traditional search (here are ten links, good luck) or AI chat (here's an answer with no citations), Perplexity gives me synthesized answers with inline citations to the actual sources. When I'm researching a new technology, checking facts for an article, or trying to understand a topic I know nothing about, Perplexity gets me oriented faster than any other tool.
The Deep Research feature takes this further. For complex research questions, it runs dozens of searches, reads hundreds of sources, and produces comprehensive reports. I've used it for competitive analysis, technical deep-dives, and understanding regulatory landscapes. The output quality varies, but when it works, it saves hours of manual research.
Perplexity has also expanded into agentic territory with Comet, their AI-native browser that can autonomously navigate the web to complete tasks. It's still early, but the vision of research that extends beyond finding information to actually doing something with it is compelling.
Where Perplexity falls short: It's not a thinking partner. Perplexity excels at finding and synthesizing information, but I wouldn't use it to analyze a business decision or get feedback on a strategy. Different tools, different jobs.
Writing code with AI assistance has fundamentally changed how I work. But the tooling here is still rough around the edges.
Claude Code lives in my terminal, which is exactly where I want it. Rather than adding another interface to manage, it enhances my existing workflow.
The core value proposition is simple: I describe what I want in natural language, and Claude Code helps me build it. But what makes it powerful isn't just code generation—it's the codebase understanding. Claude Code indexes my projects and maintains context about file structures, dependencies, and patterns. When I ask it to add a feature or fix a bug, it understands where things are.
The recent updates have been meaningful. Session management lets me name and resume conversations, so I can pick up complex work across multiple sessions. The checkpoint system automatically saves code state before each change, making it easy to rewind when things go wrong (which happens). And the slash commands system lets me create shortcuts for common operations.
I use CLAUDE.md files extensively—these provide context and instructions that persist across sessions. For each project, I maintain a file that explains the architecture, coding conventions, and common patterns. This dramatically improves the quality of suggestions.
Where Claude Code falls short: Complex multi-file refactors still require heavy supervision. Claude Code is excellent at understanding and modifying existing code, but when changes cascade across many files, things break in subtle ways. I treat it as a capable pair programmer who occasionally makes rookie mistakes.
While Claude Code handles terminal work, Cursor has become my primary IDE.
Cursor is built on VS Code but reimagined with AI as a first-class citizen. The tab completion feels eerily prescient—it predicts not just what I'm typing but what I'm trying to do. The Agent Mode can make changes across multiple files based on natural language instructions, though I find myself using this carefully (more on that in a moment).
The model flexibility is a real advantage. Cursor lets me choose between GPT-4, Claude, and other models depending on the task. Different models have different strengths, and being able to switch on the fly is valuable.
What I appreciate most is how Cursor handles context. When I ask it about code, it understands the surrounding files, the imports, the project structure. This context-awareness makes the suggestions dramatically more useful than generic code generation.
Where Cursor falls short: Agent Mode, while impressive, still requires supervision for anything non-trivial. I've had it make changes that looked correct but introduced subtle bugs. For routine tasks it's excellent; for complex work, I treat it as a first draft generator rather than a solution. The parallel agents feature—running up to eight agents simultaneously—sounds powerful but has been unreliable in practice.
Voice input has become essential to my workflow, though it took time to find the right tool.
SuperWhisper handles voice-to-text on my Mac and iPhone, and it's earned its place through reliability, privacy, and surprising versatility.
The killer feature is that it runs locally. My voice data doesn't leave my device unless I explicitly choose cloud processing. For someone who dictates notes throughout the day—often capturing half-formed ideas or sensitive business thoughts—this matters enormously.
The activation is effortless: Option+Space brings up the input. I talk. Text appears in whatever application has focus. Done. The friction is so low that I actually use it, which wasn't true of previous dictation tools.
Accuracy is genuinely impressive, even with technical jargon and unusual names. The multiple AI models (from lightweight "Nano" to full "Ultra") let me balance speed against accuracy depending on the situation. For quick messages, Nano works fine. For longer dictation, Ultra catches nuances that matter.
What sets SuperWhisper apart from basic transcription is its "Modes" feature. These AI-powered modes can automatically format and clean up your dictation—converting rambling speech into properly structured text, adjusting tone for different contexts, or formatting for specific output types. It's not just a pipe from voice to text; it can intelligently process what you say.
Where SuperWhisper falls short: While the Modes feature adds AI-powered formatting, it's still primarily a capture tool rather than an interactive assistant. If you want to have a back-and-forth conversation or get AI feedback on your dictated content, you'll need to bring that text into another tool.
The tools above need surfaces to run on. These are the environments where my work actually happens.
Warp has replaced my previous terminal application completely.
The AI integration is genuinely useful. Pressing '#' lets me describe what I want in natural language, and Warp suggests the command. Right-clicking an error and asking for explanation actually explains things. These sound like gimmicks, but they've meaningfully reduced my friction with command-line work.
The blocks concept—where each command and its output are grouped as a unit—makes navigating history dramatically easier. I can scroll back, find a specific output, and copy exactly what I need without hunting through walls of text.
Warp's agent modes go further. Pair mode suggests options while I maintain control. Dispatch mode runs autonomously, making decisions without explicit permission at each step. I use Pair mode constantly; Dispatch mode carefully.
Where Warp falls short: It's heavy. Warp uses more resources than a basic terminal, and sometimes I miss the simplicity of iTerm. The AI features require network connectivity, which occasionally matters when I'm working offline or on flaky connections.
Every tool I've mentioned produces artifacts—notes, decisions, research, code snippets. Those artifacts need to live somewhere searchable and connected.
Notion serves as my knowledge layer. Meeting notes, project documentation, personal reference material, draft articles—it all lives in Notion. The AI capabilities have made this more useful. I can ask questions across my workspace and get answers that pull from multiple pages and databases. When Notion connects to my Google Drive and Slack, those answers incorporate even more context.
The AI writing assistance has matured considerably. I use it for first drafts, for rewriting dense text, and for summarizing long documents. The ability to @mention existing pages as context means the AI writes in a style consistent with my other work.
Where Notion falls short: The AI features now require Business or Enterprise plans. This is a meaningful pricing change that's pushed some users to alternatives. The offline mode, while finally available, still feels like an afterthought compared to native apps.
Slack remains my primary work communication tool, and the AI layer on top has become genuinely useful.
Channel recaps summarize conversations I missed—essential when returning from focus time or time off. Thread summaries help me quickly understand long discussions without reading every message. Huddle transcription captures meeting notes automatically.
The third-party AI integrations extend this further. Having Claude, Perplexity, and other AI tools available directly in Slack means I can get quick answers without context-switching. When someone shares a Notion link, I get an AI summary inline.
Where Slack falls short: The AI add-on pricing has shifted around, and the feature availability varies by plan in ways that aren't always clear. The AI-generated summaries occasionally miss nuance in ways that matter. I trust them for orientation, not for detail.
Despite this stack of tools, there are real gaps.
Seamless handoffs: Moving context between tools remains manual. When I'm researching in Perplexity, reasoning in Claude, and implementing in Cursor, I'm constantly copying and pasting. Each tool understands its own context deeply but knows nothing about what happened in the other tools.
Reliable autonomy: Every tool I've described requires supervision. The "agents will do everything" vision isn't reality yet. When I let tools work autonomously, I get outputs that look correct but contain subtle errors. This isn't a complaint—I'd rather have tools that require supervision than autonomous tools that fail invisibly. But it's a gap.
Consistent memory: Some tools remember, some don't. The ones that remember sometimes forget things I wish they'd kept and keep things I'd rather they forgot. There's no unified memory layer across my stack.
Cost transparency: I pay for Claude, ChatGPT, Cursor, Warp, Notion, and others. The subscriptions add up. Worse, usage-based pricing for API features makes costs unpredictable. I'd love a single dashboard that showed me what I'm paying and what I'm getting.
Here's what I want to be honest about: this stack is amazing and frustrating in equal measure.
On my best days, I'm genuinely more capable than I was a year ago. I can research topics faster, think through problems more thoroughly, write code more confidently, and capture ideas more reliably. The tools don't make me smarter, but they let me leverage what intelligence I have more effectively.
On my worst days, I spend time fighting with tools that should work but don't. I get confidently wrong answers that waste my time. I watch tokens burn on failed attempts. I manually copy context between applications that should share it automatically.
Both experiences are true. Both are daily occurrences.
If you're building your own stack, my advice is simple: start with one tool, use it until you understand its limits, then add the next. Don't try to adopt everything at once. And maintain healthy skepticism—these tools are good but not magic.
In the next articles in this series, I'll dive deeper into specific layers of the stack. How I actually use Claude for complex reasoning. The workflow that makes Cursor productive rather than problematic. The voice capture patterns that turned SuperWhisper from novelty to necessity.
But that's the overview. This is what I actually use, every day. Not aspirational—functional. Not perfect—useful.
Discover more content:
