The AI tool landscape moves so fast that by the time a list gets published, half of it is stale. So instead of a rankings board, here is the honest account of what I keep installed and reach for when the work actually needs a tool. No affiliate links, no "disrupt 2025" energy.
Identity, Not Hype
The best way to think about these tools is by what problem they kill, not by brand. I group everything in my own workflow into four buckets.
1. For Writing Code
The fundamental trick here: you are not asking a model to write your whole app. You are asking it to stop being the bottleneck for the parts that are tedious or unfamiliar.
- Claude Code — built for long, multi-file editing sessions right in the terminal. Its superpower is finishing a chain of tasks instead of answering one question and stopping.
- GitHub Copilot — shines inside your editor for inline completions. Great for boilerplate and test scaffolds where you want low-friction, not a full conversation.
- Cursor — a whole editor threaded with AI. Good when you live in the diff and want AI to touch files you can review in place.
The common thread: they are all mediocre if you hand them a vague sentence and expect magic. They are excellent if you bring a clear spec.
Garbage in: "make a landing page"
Better in: "a 3-section landing page, dark theme,
primary color #fd6f00, links to /contact"
2. For Writing Words
For prose, the bar is different. You want a model that holds a long voice and does not collapse into marketing-speak.
I keep one notetaking canvas and one long-form assistant. The trick is never to publish the first draft — always tell the model the audience, then edit the second draft by hand. Draft one is scaffolding; draft two is yours.
3. For Research and Summaries
When I need to catch up on a spec, a blog post, or a chunk of documentation, I paste it in and ask for the decisions and the open questions — not a summary of every sentence.
That framing matters. Asking for "a summary" gives you a compression. Asking for "the decisions, the open questions, and the trade-offs" gives you a map you can actually act on.
4. The Ones That Surprised Me
Two categories overdelivered versus my expectations:
- Local models for private data. Running a small model locally for redaction and classification means sensitive text never leaves your machine. That is a selling point no API can match.
- Single-purpose micro-tools. Tiny models tuned for one job — extracting fields from receipts, transcribing calls, tagging tickets — are often wildly more reliable than asking a giant model to do the same thing within a bigger prompt.
A Decision Rule
Here is the rule I use before adopting anything new:
If it saves me a switch of context or a 30-second annoyance, I probably skip it. If it removes a whole category of work — writing boilerplate, transcribing, classifying — I keep it.
That rule filters out ninety percent of the noise. The tools that change my week are never the ones that shave ten seconds. They are the ones that delete a task outright.
What To Ignore
Ignore the "AI will replace X" articles. Ignore LMSYS leaderboard chasing — those numbers measure a sandbox, not your workload. Keep your cursor on the two questions that matter: does it delete a task I currently do, and does it make my output better without slowing me down.
Build your own list around those, and you will be better off than anyone who follows a "Top 10" post.