AI Guides › Workbench
By Nigel Guy · 5 min read
Most people start learning AI by collecting tools and prompt lists. It feels productive, and it leaves you with a folder of clever phrases and no change in how you work. The skills that carry over from one tool to the next are smaller and less glamorous, and you can practise them this week on work you already do.
A note on sourcing. This guide was built from a short social-style piece about which skills to learn first. Its claims about who is learning what, and which skill "pays the bills", are personal opinion and cannot be verified. So this guide keeps the part that holds up: three durable skills, each tied to what the vendors' own documentation says, and a way to practise each one. The order and the pay-off claims below are our judgement, not a measured result.
The rule: learn to brief the model, learn to check what comes back, and only then learn to point both at a job you are already paid to do.
| Skill | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| 1. Briefing | Gives the model the context, goal and format it needs, so the first answer is usable | £0 on a free plan | Anyone starting out | Longer is not better; vague goals still give vague answers |
| 2. Checking | Catches invented facts, wrong numbers and confident nonsense before they leave your desk | £0, just time | Anyone sending AI-assisted work to other people | Slow at first; skipped under deadline pressure |
| 3. Applying | Wires the first two into one repeated task from your own job | Free tier to start; paid plans if you hit limits | People with a recurring task | Needs a task you understand well enough to judge the output |
Cost detail: Claude's free plan is £0. Claude Pro is listed at $20 a month billed monthly, or $17 a month on annual billing ($200 upfront). That is about £15 a month on the monthly plan at time of writing; check the £ price at checkout, as exchange rates and tax vary. Other assistants have free tiers too, but we could not confirm another vendor's current plan prices for this edition, so check their pricing pages yourself.
Anthropic's own guidance says to begin with a clear definition of success and a way to test against it, then draft the prompt. Its prompting reference covers clarity, examples, structured tags, role, thinking and chaining prompts. You do not need all of that on day one. You need four things in every request: who the answer is for, what a good result looks like, what to use as source material, and what shape the output takes.
Practise on one real task, not a toy. Run this:
You are a careful editor helping me with [TASK, e.g. a client update email].
Context: [WHO THE READER IS, WHAT THEY ALREADY KNOW, AND WHAT I WANT THEM TO DO].
Source material: [PASTE THE NOTES OR FACTS TO USE; USE NOTHING ELSE].
A good result: [3 SHORT CRITERIA, e.g. under 150 words, plain tone, one clear ask].
Format: [OUTPUT SHAPE, e.g. subject line plus short body].
If anything above is missing or unclear, ask me up to three questions before you write.
Before you answer, check your draft against my criteria and fix any miss.
Fill in the square brackets. The "ask me first" line matters: it stops the model guessing at gaps you did not notice you had.
Models can state wrong things fluently. Checking is a routine, not a feeling. After any answer that will be used, run it through this three-pass check.
Act as a sceptical reviewer of the text below. Do not rewrite it.
Text: [PASTE THE AI-GENERATED ANSWER].
Original source material: [PASTE THE FACTS IT SHOULD HAVE USED].
Step 1: List every factual claim, number, name and date in the text.
Step 2: Mark each one as SUPPORTED by the source material, NOT IN SOURCE, or CONTRADICTED.
Step 3: List anything stated more confidently than the source allows.
Output a table with columns: claim, status, where in the source. Do not guess: if you cannot tell, mark it NOT IN SOURCE.
The model reviewing its own work is a first filter, not proof. You still check every NOT IN SOURCE and CONTRADICTED line against the real document, price page or person.
Pick one task that you repeat, that you know well, and that has a visible output: weekly summaries, first-draft replies, meeting notes into actions. Run the briefing prompt, then the checking prompt, and time yourself against doing it unaided. Keep the version of the prompt that worked and reuse it. That saved, tested prompt is the thing worth having, far more than a library of other people's.
I do this task regularly: [TASK]. Here is a recent real example of the input: [INPUT, WITH ANY PRIVATE DETAILS REMOVED] and the output I was happy with: [GOOD OUTPUT].
Write me a reusable prompt for this task with placeholders in [SQUARE_BRACKETS] for the parts that change each time. Include a role, the goal, the output format, three guardrails, and an instruction to ask for missing inputs. Then list two ways the prompt could go wrong.
Which skill earns money depends on your job. The sensible reading is that checking and applying are where the value sits, because anyone can get a draft; fewer people can tell whether it is right. We cannot back that with data, so treat it as a working view.