What AI I Use How and When (November 2025 Update)
This is a follow-up to my March 2025 report on the AI tools I use. Rather than exhaustively listing everything again, I’m focusing on what’s changed.
Why I Keep Updating This
AI technology develops remarkably fast. Tools evolve, new capabilities emerge, and what worked brilliantly six months ago might now be obsolete. My approach: test new things regularly when they become available, especially if I hear good things from other practitioners or read compelling coverage. I try to take these tools seriously — complete actual tasks with them, see what works best, then decide whether to adopt them properly and learn them well, or park them for later review. I also periodically reassess existing tools to see if I’m still as interested in them as I used to be.
This is what I recommend to anyone working seriously with AI: constant experimentation balanced with building deep expertise in your core tools.
The Big Change: Claude Is Now My Daily Driver
Trend: ↗️↗️↗️ From barely using it to primary chatbot and agent
In March, I wasn’t using Claude much and didn’t think much of it. That’s completely changed. Claude has become my daily driver for both conversational AI and agentic work.
Where: Web, Phone, Tablet, Terminal
How Much: Max ($100/m)
Why the switch:
The Claude 4.5 series is genuinely excellent. Previous versions felt unbalanced — good at some things, lacking in others. But Sonnet 4.5 is a superb all-around model with great style, high intelligence, extended thinking capability, and the best agentic behaviour currently availble. It takes actions reliably and effectively.
More importantly, I was hugely impressed by Claude Skills. The skills format resonates with me as a way to build on the model’s agentic capabilities. I now use skills heavily in both the web app and Claude Code to complete various tasks using scripts and instructions. I’ve even created a simple MCP server called Skillz to let other agents use skills.
What I particularly appreciate:
Style — Claude has excellent writing style and personality
Intelligence — The 4.5 series is properly smart, Sonnet for most work, Haiku for simple tasks
Deep Research — The best research functionality available from major providers, leveraging Claude’s agentic capabilities rather than bolted-on features
Projects — Excellent implementation, especially not that it uses agentic search within projects
Thinking — Unlike GPT-5 (which is terrible without thinking), Claude works quite well even without thinking enabled, and you can layer extended thinking on top for more complex work
What I Still Use (And Why)
ChatGPT
Trend: ↘️ Still using but for specific strengths
GPT-5 with extended thinking remains the most intelligent model currently available for solving genuinely hard problems. I also occasionally use GPT-5 Pro via the OpenAI API Playground when I need maximum intelligence for optimal solutions (decided not to continue the expensive ChatGPT Pro subscription since I don’t use it enough).
Key use cases:
GPT-5 with thinking for research and finding diverse sources on the internet
Hard problems requiring maximum reasoning
Critical note: GPT-5 without thinking is basically useless. If you’re disappointed by ChatGPT, always use the thinking model. Takes longer, but actually works.
Gemini
Trend: ↔️ Regular use for specific tasks
Gemini 2.5 Pro is remarkably balanced and neutral in its style. Unlike Claude (which inserts its own style) and GPT-5 (which I find difficult to control for writing), Gemini lets me set the style and define the task precisely. It’s slow — spends considerable time thinking like GPT-5 — but produces genuinely good results.
Key strengths:
Writing tasks where I need precise control
Multi-modal tasks like audio transcription and image generation
YouTube video analysis and summarisation
Native integration with Google properties
Good balance between intelligence and controllability
What I’ve Stopped Using
Perplexity
Trend: ⬇️⬇️⬇️ No longer using
I was quite excited about Perplexity a year ago, but it’s become increasingly gimmicky. They focus on dazzling features over quality. The search functionality I appreciated is now available broadly — ChatGPT and Gemini both do excellent search work, and Google AI mode is quite good at search combined with AI.
Google AI Studio and NotebookLM
Trend: ↘️ Occasional use only
Both were innovative when they launched and continue evolving, but the user experience isn’t convenient for my workflow. The functionality they provide is now available from many other products, so I’ve mostly moved on whilst occasionally checking back in.
Power Tools
Raycast
Trend: ↗️↗️↗️ Using extensively for all personal work
Where: Desktop (primarily), also Phone and Tablet
How Much: Pro + Advanced AI ($20/m)
Raycast has become essential infrastructure for my AI work. I use its snippets feature heavily to manage prompts and insert them across different apps. The clipboard history is invaluable when working with multiple AI tools. The built-in AI in advanced mode gives me access to many models, very handily available.
After being Mac-only for years, it’s now available on Windows — worth checking out regardless of your platform. I recorded a video demo that got substantial attention and helped many people understand how to integrate it into their workflow.
Why it matters: Even without its AI features, Raycast is essential when you work a lot with AI. Clipboard history for moving text between tools, snippets for prompt templates, keyboard shortcuts for all your AI apps — it’s the personal infrastructure layer that makes everything else work smoothly.
Knowledge Management
Notion
Trend: ↗️↗️ Dramatically increased usage since Notion Agent launched
Where: Web, Phone, Tablet
How Much: Business ($20/m)
Notion’s AI integration was already good, but the introduction of Notion Agent a few months ago transformed it into an incredibly powerful product. The agent handles complex tasks within my workspace — updating databases, writing content, revising documents — with excellent reliability.
Key insight: Invest heavily in custom instructions for your agent and in crafting good prompts for agentic tasks. This pays off enormously. I recommend using Claude 4.5 Sonnet as your model choice in Notion Agent — it’s excellent at agentic work and better at writing, which is primarily what we do in Notion.
Readwise Reader
Trend: ↔️ Still using regularly but relying less on its AI features
Where: Web, Phone, Tablet
How Much: Full ($8/m)
I continue using Reader for consuming content, but integration with other AI tools has become so seamless that I’m relying less on Ghost Reader (the built-in AI). Instead, I use Raycast, Claude, and other products to process information that comes through Readwise.
Coding Tools
Trend: ↗️↗️↗️ Agentic coding has become a significant occupation
I’ve done extensive research in this area and am teaching a course with Isaac Flath on agentic coding. We’ve tested many tools and documented them on our AI Coding Substack.
Visual Studio Code with GitHub Copilot
Trend: ↔️ Primary software development setup
Where: Desktop, Web, CI
How Much: GitHub Copilot Pro+ ($39/m) — disclosure: my subscription is generously sponsored by GitHub as part of their program for community contributors
This remains my main coding environment. Works excellently, I particularly appreciate that these components are part of open-source VS Code. GitHub Copilot gives me access to many different models — all the latest from OpenAI, Anthropic, plus various open and experimental models — all integrated smoothly.
Usage patterns:
Primarily use agent mode (I’ve become quite skilled at agentic coding)
Still use inline editing regularly
Tab completion, which was the original way to do AI coding in IDEs, I use less and less — I simply don’t write that much code directly anymore in ways that completion would help
Claude Code
Trend: ↗️↗️Primary terminal-based agent
Where: Terminal, Web
How Much: Included with Claude Max (I never reach the limits, but relieved that Anthropic now provides a way to pay-per-consumption for overages)
Definitely the best product in the terminal-based agent space. Very versatile. The limitation to Anthropic models only is manageable since the Claude 4.5 series is good enough for many tasks.
For tasks requiring more intelligence — analysis and planning — I also use Codex from OpenAI. GPT-5 is definitely the more intelligent model. I’ve written about integrating GPT-5 with Claude Code for this purpose. Occasionally I also use Gemini CLI (for multi-modal tasks) and OpenCode (for using a variety of models).
Remote Asynchronous Agents
Trend: ⬆️⬆️⬆️ Increasingly central to my workflow
These are the best way to work productively with AI coding tools without directly babysitting them whilst they work.
Main tools:
GitHub Copilot Coding Agent — Very easy to use, available on GitHub. Just assign issues or trigger tasks, including from within VS Code
Codex Cloud — Excellent tool for async work
Claude Code Web — Also very good for remote tasks
Much of my coding work now happens by triggering tasks, assigning them to an agent, and letting them work in the background.
Warp Terminal
Trend: ↗️ Daily use
Where: Terminal
How Much: Warp Build ($20/m — pay-per-consumption and BYOK available for overages)
Excellent terminal with great integrated AI capabilities. I don’t use it for software development projects per se (though it’s perfectly adequate), but rather for executing quick tasks — writing ad hoc scripts, handling terminal operations I don’t remember exactly how to do. Excellent work integrating AI into a terminal, and I enjoy using it.
What I’ve Stopped Using
“Vibe coding” tools like v0 — I enjoyed experimenting with them, but they’re no longer necessary for my work. I can now do agentic coding with standard tooling.
Audio
Monologue (replacing MacWhisper)
Trend: ⬆️⬆️ Switched completely, using all day
Where: Desktop
How Much: Monologue Pro ($10/m)
I rely heavily on dictation for my work. Previously used MacWhisper, which is fine, but discovered Monologue — a really nice tool with excellent transcription, very nice interaction mode, easy customisation, and cloud-based transcription (no need to run models locally).
I use this all day long to talk to the computer. Fundamental to how I work.
Speechify
Trend: ⬆️ New addition
Where: Browser, OS
How Much: Premium ($12/m)
Speechify does the opposite of Monologue — text-to-speech. The voices aren’t quite as good as Eleven Labs or OpenAI’s offerings (slightly robotic), but the product integration is superb. Works everywhere I work — chatbots, web pages, web applications — through excellent browser and OS integration.
From App to Skill: CardCraft
One of the first casualties when I discovered Claude Skills was my custom app, CardCraft, for creating flashcards. It’s now available as a skill I run within Claude or Claude Code, and it works just as well.
This is an excellent example of taking functionality and packaging it as a skill instead of maintaining a dedicated app. I wrote about this transition: Are Skills the New Apps?
The skill is available as a GitHub Gist if you want to try it.
Final Thoughts
The tools landscape continues evolving rapidly. The major theme for me has been the shift from multiple specialised tools to more capable, genuinely agentic systems that can handle diverse tasks reliably. I continue to experiment and try new tools and techniques and hope to write an updated report when my toolbox changes again in a few months.