AI Guides › Workbench
By Nigel Guy · 7 min read
Most people collect AI tools the way they collect browser tabs. Each one gets opened for a week, none gets set up properly, and every job starts with the question "which one should I use for this?" That question is the cost. A tool you half-use never learns your voice, never holds your files and never becomes a habit.
The rule: every recurring job gets exactly one named tool, and each tool gets set up once so it stops asking you to explain yourself.
This guide gives you a ledger to fill in, a starter set of five jobs with tools we can verify today, and a ready-to-adapt email skill. The picks are a sensible default, not a league table. Swap any of them if you already pay for something that does the job.
A ledger is just a table with four columns: the job, the one tool that owns it, the one-time setup, and the check that tells you it is working. If a job has two owners, it has none.
| Job | Owner | One-time setup | Check |
|---|---|---|---|
| Writing that sounds like you | Claude | A custom Skill holding your rules | Draft needs fewer than two edits |
| Questions about your own documents | Gemini Notebook | One notebook per topic, sources added | Answers cite your sources |
| Live-web research with sources | ChatGPT (search, deep research) | None; paste your question template | Every claim has a link you opened |
| Repeating tasks on a timer | ChatGPT (scheduled tasks) | One task, one clear prompt | It runs without you |
| Building and fixing code or files | Claude Code | Project folder and a short instruction file | It runs the tests, you read the diff |
Prices are in US dollars on the vendor pages, with approximate pounds. Check the £ price at checkout, because tax and exchange rates change.
| Tool | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Claude | Chat, Skills, files, Claude Code | Free; Pro $20 a month (about £15 at time of writing, or $17 a month billed annually); Max from $100 a month (about £75 at time of writing) | Writing, long documents, code | Heavier use runs into plan limits |
| ChatGPT | Chat, web search, deep research, projects, scheduled tasks | Free, Go, Plus and Pro tiers; the pricing page I could read showed no figures, so check the £ price at checkout | General all-rounder, scheduled tasks | Go and Plus carry different limits; Go may include ads per OpenAI |
| Gemini Notebook | Research workspace over sources you supply | Pricing not confirmed in the pages I could read; check the current page | Studying your own PDFs and notes | See the naming note below |
A naming note: Google's NotebookLM address now redirects to a product called Gemini Notebook, described as an "AI Research Tool & Thinking Partner". I could not confirm whether this is a rename or a successor, or what the plans cost. Check Google's own page before you pay.
The usual failure is pasting "write in a friendly tone" into every chat. Put your rules in a Skill instead, and Claude loads them only when the task calls for them.
Skills are folders of instructions Claude loads when relevant. Anthropic's help centre lists them as available on Free, Pro, Max, Team and Enterprise plans, with code execution enabled. To set one up:
The description matters most. Claude uses it to decide whether the skill applies, so say what it does and when: "Drafts and replies to email in my house style. Use for any email request."
Here is a body for the email skill. Replace everything in square brackets.
You are my email drafting assistant. Your job is to write emails
that read as if I wrote them on a good day.
VOICE
- [3 to 5 rules, for example: British English, warm but brief,
contractions allowed, one clear request per email]
SHAPE
- Lead with the point; no opening pleasantry unless the thread needs one.
- [Your length, greeting and bullet-or-prose rules]
BANNED
- Never use: [phrases you hate, for example "touching base",
"hope you are well", exclamation marks]
CLOSE
- End exactly with: [your sign-off and name block]
HOW TO WORK
1. If I have not given you the recipient, the goal or the thread
I am replying to, ask for them. Do not guess.
2. Match the length of the message I am replying to.
3. When the tone is a judgement call, give two versions and say
which you would send.
4. After the draft, list anything you assumed rather than knew.
5. Before answering, check the draft against VOICE, BANNED and
CLOSE and fix any breach silently.
Fill in: your voice rules, banned phrases and sign-off. Paste three emails you were pleased with into the chat on the first run so the rules have examples behind them.
Create one notebook per topic, add the sources, and ask questions of those sources only. This beats pasting documents into a general chat because the answers stay tied to what you supplied. Test it by asking a question whose answer you already know and confirming the citation points to the right passage.
OpenAI's pricing page lists search on every plan and deep research on every plan, limited on Free and Go and fuller on Plus and Pro. Use this prompt template.
You are a careful research assistant. I need an answer I can rely on.
QUESTION: [YOUR_QUESTION]
AUDIENCE AND USE: [WHO_WILL_READ_THIS_AND_WHAT_DECISION_IT_FEEDS]
REGION AND DATE: [COUNTRY], current as of [DATE]
Method:
1. Prefer official documentation, regulators and primary data over blogs.
2. Give a short answer first, then the evidence as a table:
claim, source, date of source.
3. Mark anything you could not confirm as UNVERIFIED.
4. Do not fill gaps with plausible-sounding figures. If a number is
missing, say so.
5. If the question is ambiguous, ask me one clarifying question first.
Before you answer, check that every claim in the table has a source.
Fill in: the question, the audience and the region. Then open at least the three links that matter most.
The pricing page lists scheduled tasks on Go, Plus and Pro but not Free. Use them for jobs you would otherwise forget, such as a Monday summary of a topic you track. Write the task prompt as if the reader has never seen your chat, because on the day it runs, that is true.
Anthropic's pricing page lists Claude Code under Pro. Keep a short instruction file in the project folder, ask for small changes, and read every diff before you accept it.