AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
The usual way to "keep up" with AI is to learn whatever launched this week. It feels productive because there is always a new tool, a new tutorial and a new thing to be behind on. It fails because tool knowledge goes stale faster than you can use it, and nothing you learn is tied to work you actually do.
The rule: only build a skill that survives a tool change, that you will use on real work within a fortnight, and that you can check for yourself.
This guide gives you a test to run first, then five skill areas that pass it for most people with a job or a small business. You do not need all five. You need one, chosen on purpose.
Take each skill you are tempted to learn and answer three questions in writing.
| Question | Pass | Fail |
|---|---|---|
| Does it outlive the tool? | The skill still applies if you swap one assistant for another | It is a menu path or a feature name |
| Is there a task waiting? | You can name a real task from the last two weeks it would improve | You would be "exploring" |
| Can you check the result? | You can tell good output from bad without asking the AI | You would have to trust it |
Two passes out of three is a maybe. Fewer is a skip. The third question matters most: a skill you cannot verify only produces confident mistakes faster.
Anthropic describes context engineering as curating the set of information the model sees, and notes that performance can degrade as the context grows ("context rot"). In practice: give the model the few things it needs, not everything you have. Practise by rewriting one recurring request as role, goal, inputs, constraints and format. Passes all three questions for almost everyone.
Anthropic's guidance on testing says success criteria should be specific, measurable, achievable and relevant, and that you should test on cases that mirror real work, including awkward ones. For you, that means: before using AI on a recurring task, write five example inputs and what a correct answer looks like. Run them each time you change the prompt or the tool. Where you use AI to grade AI, Anthropic's advice is to use a different model for grading than for generating.
If you do the same task weekly, turn it into a reusable instruction. In Claude Code this is a skill: a SKILL.md file, in ~/.claude/skills/<skill-name>/ for personal use or .claude/skills/<skill-name>/ in a project, with a description that tells Claude when to use it. Other assistants have equivalents (saved instructions, custom assistants). The durable skill is writing down a process clearly enough that someone else could follow it. Skip this until you have done the task by hand at least three times.
Pick the claims, numbers and names in an output that would cost you something if wrong, and check those against a primary source. This is the unglamorous centre of the whole list. The World Economic Forum's Future of Jobs Report 2025 reportedly puts AI and big data at the top of its fastest-growing skills and analytical thinking as the most sought-after core skill; I could not open the report page directly when checking this guide, so treat that as a pointer to read it yourself rather than a figure to quote.
If you handle customer or staff information in the UK, data protection law applies to how AI tools process it. The ICO publishes guidance on AI and data protection for organisations covering accountability, transparency, lawfulness, accuracy and fairness. The skill is a habit: before pasting anything, ask whether it identifies a person, and whether your plan's terms allow it. Check your tool's current data settings; they differ by plan and change.
Score each skill using the test, choose the highest, and write one line: "By [DATE] I will have used [SKILL] on [REAL TASK] and checked the result by [METHOD]." If you cannot fill all three, you picked an interest, not a skill.
A prompt to make the test sharper:
Act as a sceptical learning coach. I want to decide which AI skill to
build next. My role: [ROLE]. My recurring tasks: [LIST 3-5 TASKS WITH
HOW OFTEN]. Skills I'm considering: [LIST].
For each skill, ask me the three questions: does it outlive the tool,
is there a real task from the last two weeks, can I check the result
myself. If any of my inputs are missing or vague, ask me before
scoring. Do not assume details about my work.
Output a table (skill, three yes/no/unclear answers, one-line reason),
then recommend exactly one skill and one first exercise under 60
minutes. Before answering, check that every answer cites something I
told you, not something you guessed. Say plainly if I should skip all
of them.
Fill in your role, tasks and candidate skills.