AI Guides › Playbooks
By Nigel Guy · 8 min read
Most people decide whose AI advice to follow the way they'd pick a queue at the supermarket: by which one looks most confident. A tidy thread, a big follower count, an impressive screenshot. None of those tell you whether the person has actually done the thing they're describing, or just describes it well. Fluent explanation is cheap now. The tools you're getting advice about will produce it on request.
The rule: before you act on anyone's AI advice, ask four questions about the person giving it, and let the answers set how much checking you do yourself. That includes the advice on this site.
A receipt is dated evidence that something happened. Each question gets a pass, partial or fail. You don't need four passes; you need to know which answers are missing, so you know what to verify yourself.
| # | Question | What a pass looks like | What a fail looks like |
|---|---|---|---|
| 1 | What did you actually do, and can I see it? | A specific setup, file, prompt or result you can inspect or repeat | "I've helped hundreds of businesses", with no single example you can look at |
| 2 | When did you last check it? | A date, a version or plan name, and the menu path or command as it is now | No date, or screenshots of an interface that has since changed |
| 3 | What do you get if I follow it? | Affiliate links, sponsorships and their own products stated plainly | You find out about the course or the referral code only after you've clicked |
| 4 | Where did it go wrong for you? | Named limits, edge cases and failures from their own use | Every outcome is positive and nothing ever broke |
Someone who has done the thing can point at it: the exact prompt, the settings, the plan they were on, the output before they tidied it. Talking about a category ("agents will change your workflow") is commentary. One workflow you can rebuild is experience. A real example with client details removed is a fair partial pass.
AI products change names, plans, limits and menus often, so advice that was right in spring can be wrong by autumn. A dated claim ("as of September, on the paid plan, Settings then Connectors") can be checked against the vendor's help page in minutes. If there's no date, treat every price, limit and menu name as unverified until you've found it on the vendor's own page.
Being paid isn't disqualifying. What matters is whether they told you. In the UK, the advertising regulator's guidance for influencers expects ads to be obvious without you having to click or interact to find out, with "#ad" as the minimum label, and it covers affiliate links too. Since 6 April 2025, the Digital Markets, Competition and Consumers Act has also made it unlawful to post or commission fake reviews, or reviews that hide that they were incentivised. If you only discover the commercial angle by digging, take that as your answer to Question 3.
This is the hardest one to fake. People who use a tool in real work know where it breaks: the file size it chokes on, the task it gets confidently wrong, the setting that quietly resets. If someone can't name a single limitation, they either haven't used it much or aren't telling you.
The check only means something if it applies to us as well. Test every line of this yourself.
| # | Question | Our answer, and how you can check it |
|---|---|---|
| 1 | What did you do? | Guides give the exact clicks, settings and prompts so you can repeat them. If a step can't be reproduced from the guide alone, that's a failure on our part. |
| 2 | When did you check? | Every guide has a "Last verified" date at the top and a list of sources at the bottom, mostly vendors' own documentation. If the date is old, check the sources before trusting the details. |
| 3 | What do we get? | Juliemango is a business selling AI tools, so we have a commercial interest in you using AI at all. Weigh our enthusiasm with that in mind. |
| 4 | Where did it go wrong? | Every guide ends with "What to skip" and "Guardrails". If a guide has nothing to say there, it isn't finished. |
Most of what we publish is basic on purpose. At beginner level you can test the advice yourself, often in minutes on a free or entry-level plan, and it's where bad advice costs most people: a wrong setting, an overshared file, a subscription you didn't need. Advanced claims often depend on a setup or budget you can't see, which makes Questions 1 and 4 hard to answer honestly. Where we go further, we try to say so and show the working.
Hypothetical scenario. A post in your feed says: "Stop paying for three AI subscriptions. One tool now does research, writing and spreadsheets — here's my exact setup," followed by a link.
That doesn't prove the advice wrong. It means all the checking is yours: read the tool's own plan page, try the free tier on one real task, and only then think about cancelling anything.
AI can't tell you whether a stranger is honest, but it can separate checkable claims from ones you'd take on trust. Fill in the post and the decision it's pushing you towards.
You are helping me judge a piece of AI-tool advice before I act on it. You cannot verify who the author is or what they have done, so do not guess about them.
The advice (pasted exactly):
[PASTE_THE_POST_OR_ADVICE]
What it is pushing me to do:
[THE_DECISION_E.G._CANCEL_A_SUBSCRIPTION_OR_BUY_A_COURSE]
Steps:
1. List each specific factual claim (prices, plan names, limits, features, menu paths, results). Quote the words each claim relies on.
2. For each claim, say where I could check it, such as the vendor's pricing page, help centre or release notes, and whether the claim has a date or version attached.
3. Answer these four questions using only what is in the text: (a) Is there something specific I could inspect or repeat? (b) Is it dated? (c) Is any payment, affiliate link or product of the author's disclosed? (d) Does it name any limitations or failures? Mark each Pass, Partial or Fail, with a one-line reason.
4. List what is missing that I would need before acting.
Output: a table of claims (claim | where to check | dated?), then the four answers, then the missing-information list. No overall verdict on whether the author is trustworthy.
Rules: do not state current prices or features from memory as fact; label anything you recall as "check this". If the pasted text is incomplete or my decision is unclear, ask me before answering.
Before you reply, check that every claim in your table appears in the pasted text and that you have not added any.
The trap is using the check to decide whether advice is true. It doesn't do that. Someone with no receipts can be right, and someone who passes all four can still be wrong this month because the product changed under them. The check tells you how much verifying you need to do yourself, nothing more.
The second trap is accepting things that only look like receipts. Follower counts, "as seen in" logos and a confident tone answer none of the four questions. A screenshot is a partial pass at most: it shows a result, not what it took or how often it worked.