There are roughly a thousand AI tools shipping in any given week. You do not have time to evaluate them. We do not have time to evaluate them. The honest answer to "which AI tools should I use" is: pick a small set, learn them well, ignore the rest.
This post is the small set. Specifically, the tools we have in our daily flow at Moffhub right now. Not what is best on paper. What survives a year of heavy use.
The list will be wrong by the time you read it. The categories will not.
One general-purpose chat product
You need exactly one. Switching between them costs more attention than the marginal capability difference is worth. Pick whichever you find pleasant to talk to and stop.
Claude (Anthropic). Our default. Best for long-form writing, code review, document analysis, and any task where you want the model to push back on a bad idea instead of agreeing. The 1M-token context window means you can paste a whole codebase or a 200-page contract and it actually engages with the whole thing.
ChatGPT (OpenAI). Strongest at math-flavoured reasoning, image generation, and tool use. The "advanced data analysis" mode is the easiest way to do real data work without writing code. Chat memory and Custom GPTs are nicer here than the equivalents elsewhere.
Gemini (Google). Tightly integrated with Google Workspace, which matters if your company runs on Gmail and Docs. Strong at video and image understanding. Search-grounded answers come for free.
Pick one. Use it for at least a month before you decide it is wrong for you. Most people who switch tools every week are running from a prompting problem, not a model problem.
One coding assistant (if you write code)
Three real options. They are increasingly similar in capability; the differences are in how they slot into your workflow.
Claude Code. Anthropic's CLI agent. Lives in your terminal, edits files in place, runs tests, iterates. Best for "given this repo, make this change". Treats the whole codebase as context, which is the right shape for most refactor work.
Cursor. A VS Code fork with deep AI integration. Tab-completion knows your project. Inline edits where you describe a change. Best for editor-resident, line-by-line work.
GitHub Copilot. The veteran. Tight integration with VS Code and JetBrains, broad enterprise support. The autocomplete is great. The chat features are weaker than the others.
We use Claude Code daily for refactors and Cursor for new code. Copilot is fine; the others are stronger now.
One research tool
When you need synthesised information with citations, not just a model's parametric guess, this category replaces a lot of search-engine work.
Perplexity. Search-first, citation-first. You ask a question; it searches, reads, and answers with footnotes. Pro tier is worth it if you research a lot.
ChatGPT or Claude with web search on. Both have built-in search now. Slightly less polished than Perplexity but you do not need a separate subscription.
For occasional research, the chat product you already use is enough. For heavy research, Perplexity earns its monthly fee.
One code-review and PR tool
If you write code, you should be using AI on every PR. There are ways to integrate this into your workflow that do not require a paid product:
ChatGPT or Claude with the diff pasted in. Free, takes 30 seconds. Works embarrassingly well for "what bugs would I miss in this diff".
CodeRabbit. Automated PR reviews on every commit. Catches the simple stuff. Mixed feelings about the more opinionated suggestions.
Greptile. Codebase-aware code review. Better than the generic options because it knows your conventions.
We mostly do option 1. The occasional Cursor / Claude Code review of a finished change catches what the human review misses.
Image and video, if you need it
Most marketing-site work, slide decks, and one-off graphics can be done in here without a designer:
ChatGPT 4o image generation. In-line image generation in the chat. Best for iterative work where you describe what you want and it adjusts.
Midjourney. Highest aesthetic ceiling. Steeper learning curve. Worth it if visual quality matters for your work.
Runway / Sora. Video generation. Still rough but improving fast. Useful for short B-roll, not for hero shots.
For most teams, ChatGPT image generation is enough. You do not need three tools here.
Voice / transcription
Underrated category. The fastest way to turn raw thought into usable document is to talk for two minutes and have AI structure the result.
Whisper (the API or any wrapper). Best transcription quality, low cost. We use this for meeting notes, voice memos, and quick dictation.
Otter.ai. Live transcription during meetings, plus AI-generated summaries. Useful if you do a lot of stand-ups or interviews.
Built-in dictation (macOS, Windows, mobile). Surprisingly good. Free. Pair with a chat product to clean up the result.
What we do not use
In the spirit of "list what you skip":
- Multi-tool aggregators that promise to give you all the models in one interface. The chat experience is always slightly worse than the native one. The novelty wears off.
- Niche AI vertical tools (AI-for-meetings, AI-for-CRM, AI-for-marketing-emails) that are mostly thin wrappers around GPT-4. The chat product can do the same job once you know how to prompt.
- AI search engines other than Perplexity. They keep launching, they keep being slightly worse than Perplexity.
- AI for tasks where AI does not help. Calendar scheduling, project management, the actual writing of code more complex than a few hundred lines. Either humans or focused tools beat AI here.
Choosing for your work, not ours
The list above is what works for a small engineering team. If you are in a different role, the shape changes:
- Lawyer or contract reviewer. Lean into Claude (long-context document work) plus a domain-specific tool like Harvey or CoCounsel for citations.
- Product manager. Chat product plus Notion AI plus a transcription tool. Skip the coding assistants.
- Marketer / writer. Chat product plus a writing-focused tool like Jasper if you do volume. Image generation matters more than coding.
- Researcher. Chat product plus Perplexity plus Elicit (academic search). Skip everything else.
The framework: pick one tool per category that genuinely matches your work. Use it for a month. Decide if it earns its place. Drop it ruthlessly if not.
How to test a tool
Before committing to a new tool, run the same five tasks through it that you currently use AI for. If three of them go better than your current tool, switch. If four of them go the same, do not. The drag of switching is real.
This stop on the roadmap is "settle on a small toolkit and stop chasing every launch". Once you have settled, your job becomes getting better at the tools you have, not finding new ones. That is where compounding starts.
Next stop: how to actually use AI in your daily work, with concrete patterns for non-coders and coders.