AI Guides › Playbooks
By Nigel Guy · 8 min read
Most beginners learn AI by collecting tools: a chatbot, an image app, a notetaker, a "prompt pack", a video generator. It feels like progress because each new tool gives a small hit of novelty. A month later you have a dozen logins and no skill you could describe, because tools change every few months and the habits underneath them never got built.
The rule: climb one layer at a time, stay on a layer until you can pass its check, and do not add a new tool until the layer you are on needs one.
This is JulieMango's own ladder, built from how the official vendor docs themselves sequence things (use it, describe it well, test it, connect it, repeat it). It is not a standard curriculum and nobody certifies it. Where we state a product fact, it is from the vendor's own pages, checked on 2026-10-04.
Each layer has one job, one check you can pass in an evening, and one thing it unlocks. Time estimates are deliberately missing: they vary too much to be honest.
| Layer | Skill | Pass this check before moving on |
|---|---|---|
| 1 | Use one assistant on real work | You have used one assistant on a real task on five different days |
| 2 | Brief it properly | Your first answer is usable more often than not, because you gave goal, context and format |
| 3 | Check its output | You can name what you verified in your last three outputs, and how |
| 4 | Iterate and show examples | You fix a weak answer with one targeted follow-up, not a restart |
| 5 | Keep reusable setups | Your three most repeated tasks each have a saved prompt or project |
| 6 | Connect it to your material | It works from your files or apps, and you know what it can see |
| 7 | Chain it into a routine | One weekly task runs from a saved process with you reviewing the result |
| 8 | Measure it | You can say what "good" means for a task and test a prompt against it |
These are the foundation. Everything later assumes them, and most people who feel "bad at AI" are missing one of these, not an advanced technique.
Layer 1: one assistant, real work. Pick one assistant and use it on work you would do anyway: an email, a summary of a long document, a plan for a trip. At time of writing, Claude's Free plan includes chat, web search, file creation and code execution, so you can start without paying. If you later want more usage, Claude Pro is listed at $20 a month (about £15 at time of writing; check the £ price at checkout), and ChatGPT Plus is listed at $20 a month, with a cheaper Go tier at $8 (about £6 at time of writing). Prices and what each tier includes change, so check the vendor's page. You do not need a paid plan to finish layers one to four.
Layer 2: brief it. Anthropic's own prompting guidance asks you to start from a clear definition of success. Translate that to everyday use: say what you want, who it is for, what you are giving it, and what shape the answer should take. A one-line request gets a generic answer.
Layer 3: check it. The assistant is fluent whether or not it is right. For each output, decide in advance what a wrong answer would look like (a made-up figure, an invented source, a wrong name) and check exactly that. Opening the source it cited is a check. Reading it again and thinking "sounds right" is not.
Layer 4: iterate and show. When an answer is weak, say what is wrong with it ("too formal, drop the second paragraph, keep the numbers") rather than regenerating. Paste an example of the output you want. An example usually does more than another paragraph of instructions.
A prompt that practises layers 2 and 3 together:
You are a careful editor helping me with [TASK, e.g. "a customer email"].
Goal: [WHAT_GOOD_LOOKS_LIKE]
Audience: [WHO_READS_IT]
Material to use: [PASTE_SOURCE_TEXT]
Format: [LENGTH_AND_STRUCTURE]
Rules:
- Use only the material I pasted for facts. If something you need is missing, ask me before writing.
- Do not invent names, numbers or quotes.
Before you answer, check your draft against the goal and the rules, then show me:
1. The draft.
2. A short list of every factual claim in it, each tagged "from my material" or "not from my material".
Fill in the square-bracket parts; the numbered list at the end is your layer 3 check done for you, but you still verify the "not from my material" items.
These turn occasional use into a system. They are worth it only once the foundation holds.
Layer 5: reusable setups. When you have typed the same brief three times, save it. Claude has Projects (listed on the Pro plan at time of writing) and other assistants have equivalents. In Claude Code, a CLAUDE.md file holds standing instructions the tool reads at the start of every session. The principle is the same everywhere: write the brief once, in a place the assistant reads automatically.
Layer 6: connect it. The Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems such as files, databases and tools, and both Claude and ChatGPT support it per the MCP site. The beginner trap here is permissions. Before connecting anything, list what the connection can read and what it can change, and start with read-only access to one folder or one app.
Layer 7: chain it. Take one weekly task and break it into steps: collect, summarise, draft, review. Claude Code's docs describe skills (packaged repeatable workflows) and scheduled tasks. You do not need to code to practise this: a saved checklist of prompts you run in order is a layer 7 habit. Keep yourself as the reviewer in the last step.
Layer 8: measure it. Anthropic's prompt engineering overview states that you should have a clear definition of success criteria and some way to test against them before you tune prompts. Build a tiny test set: five real inputs and a note of what a good answer looks like for each. Change your prompt, rerun the five, and see whether it genuinely improved. This is the layer that separates "it feels better" from "it is better".
You are helping me build a small test set for an AI task.
Task: [DESCRIBE_THE_REPEATED_TASK]
Who uses the output: [AUDIENCE]
Three real examples of inputs I have handled before: [PASTE_INPUTS]
Do this in order:
1. Ask me up to five questions about what a good result looks like. Wait for my answers.
2. Propose five test inputs (reuse mine, add two awkward edge cases).
3. For each, write a pass/fail checklist of three to five yes/no criteria.
4. Output a table: input, criteria, what a failure would look like.
If anything is ambiguous, ask rather than assume. Check that every criterion can be answered yes or no before you show me the table.
Fill in the task and paste real inputs; remove anything confidential first.
Judgement about when not to use it. It is not a layer because it runs alongside all the others: for each task, ask whether the cost of a confident wrong answer is higher than the time saved. Anything involving legal, medical, financial or safety decisions, or personal data you are not allowed to share, needs a human professional or a vetted tool, whatever the assistant says. Keep a one-line log ("used it for X, saved time, had to fix Y"). After a month it tells you where AI earns its place in your work, which no course can.
Skipping up the ladder because the upper layers are more exciting. Connectors and automations are easy to set up and hard to supervise. If layer 3 is weak, automation just produces unchecked output faster. The second trap is staying on layer 1 for ever while buying new tools. If you have a new tool in mind, ask which layer it serves and whether you have passed the layer below.
It will not make you an AI engineer, and it will not tell you which assistant is "best" this month: model lineups and plans move too fast for a guide to settle that. It also cannot replace training your employer or regulator requires. For the technical route, the free Claude Academy courses listed in the Claude Code docs are one official starting point.