AI Guides › Workbench
By Nigel Guy · 4 min read
Most people build an AI toolkit by collecting. A new app every fortnight, six tabs open, and no clear answer to "which one do I use for this?" It feels like keeping up, and it is mostly overhead. A short list you actually open beats a long list you have heard of.
A note on this guide. It started from a post about "the three tools I use", which described one person's desktop. We could not verify that person's list, and a stranger's setup is a poor template anyway, so this guide does not reproduce it. What follows is the durable part: how to pick three tools by job, and which current products fill each slot.
The rule: give every tool one job, cap the list at three, and replace a tool only when it fails its job twice.
| Slot | Job | Example tool | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|---|
| Thinker | Drafting, reasoning, code, back-and-forth | Claude | Free tier; Pro from $20 a month, or $17 a month billed annually (about £15 and £13 at time of writing); Max from $100 (about £75). Check the £ price at checkout | Writing and thinking with your own files and context | It answers from general knowledge unless you give it sources, so it can be confidently wrong on specifics |
| Reader | Answering only from documents you supply | Google's notebook tool, branded NotebookLM and now appearing as "Gemini Notebook" | Free tier exists; paid tiers sit inside Google's AI plans, so check the £ price at checkout | Long reports, transcripts, policy documents | Limited to what you upload; limits change |
| Notebook | Your own record of decisions and prompts | A plain text or Markdown file | Free | Keeping the prompts and rules that work | Only works if you write things down |
You may prefer a different product for a slot. ChatGPT, for one, fills the Thinker slot well; we could not load OpenAI's pricing page when checking, so we state no price for it. The slots matter more than the brands.
Give it a role, the goal, your inputs and a format, then make it check itself.
You are a careful editor helping me with [TASK, e.g. a client update].
Goal: [WHAT A GOOD RESULT LOOKS LIKE].
Audience: [WHO WILL READ IT].
Material: [PASTE YOUR NOTES OR DRAFT].
Steps: 1) List anything missing from my material and ask me for it before writing.
2) Draft the result in [FORMAT, e.g. 150 words, three bullets].
3) Check your draft against my material and mark any claim not supported by it.
Do not invent facts, figures or quotes. If unsure, say so.
Fill in the square-bracket parts. Step 1 is the important one: a model that asks before writing guesses less.
Upload the documents first, then ask it to stick to them.
Using only the sources in this notebook, answer: [QUESTION].
Give the answer in [FORMAT, e.g. five bullets], with the source named beside each point.
Then list anything the question needs that the sources do not cover.
If the sources disagree, show both versions rather than choosing one.
Replace the question and format. Always click through to the cited passage for anything you will act on.
One file, three headings: "Prompts that worked", "Rules I always add", "Mistakes it made". Paste in a prompt only after it has worked twice. This file is what makes a tool swap painless, because your method lives in your file rather than inside any one product.
If you are still choosing, run this over a week rather than a weekend.
I want to choose one AI tool for this job: [JOB, in one sentence].
My constraints: budget [£ PER MONTH], data sensitivity [LOW / MEDIUM / HIGH], device [DEVICE].
List up to three candidates. For each, give what it does, whether it has a free tier, and the main drawback.
Say which claims you are unsure of and tell me where to verify them on the vendor's own site.
Do not state a price unless you are confident it is current.
Fill in the job and constraints, then check every price on the vendor's page yourself.