Twenty-Nine Commands Produce Output. The Thirtieth Checks It.
Thirty commands, each one three to six words long. “Summarize this in 5 bullets.” “Find the mistakes.” “Make this shorter.” The card’s promise is in its subtitle: Copy · Paste · Get Better Results.
Twenty-nine of them do the same kind of thing. They are instructions about your material – do this to my text, my code, my topic, my notes. Explain it, shorten it, table it, quiz me on it.
Number thirty is not like that.
30. Check your answer first — Verify your response for accuracy.
That is an instruction about the model’s own output. It is the only self-checking command on the card, and it is printed last, as item 30 of 30, in a box exactly the same size as “Give me 10 captions”.
It should not be item 30. It should be a suffix welded onto about a dozen of the other twenty-nine – the ones that can hand you something false without looking any different from the ones that cannot.
And the card already knows which ones those are, because its footer names them: “Always check AI answers before you use them, especially facts, sources and code.” Three risk categories, and the card contains commands in all three.
The card this post is about
Watch the original reel: “Claude AI – 30 secret commands”
Footed “Always check AI answers before you use them, especially facts, sources and code. Results vary, so keep your own judgement.” The tables below reproduce all thirty as printed, read on 12 October 2026.
Four kinds of wrong, and only one of them announces itself
Command 30 is a good instruction. It is also a blunt one – “verify your response for accuracy” does not say against what. These thirty commands can go wrong in four distinct ways, and asking the model to check itself helps with some of them and not at all with others.
Diagram: five failure modes, mapped to commands
| Kind of wrong | Which commands produce it | What it looks like | Does command 30 catch it? | What does |
|---|---|---|---|---|
| 1. Compression loss something true was dropped |
2 summarize · 6 key points · 17 notes · 19 shorter · 16 flashcards | A clean summary that quietly omits the caveat that mattered | Partly. It can re-read the source and spot a gap | You, re-reading the original with the summary beside it |
| 2. An invented fact | 7 real example · 13 act as an expert · 18 give me an example · 9 compare A vs B | A specific, plausible, confidently stated claim | Weakly. The same model that produced it is grading it | An outside source. This is the one that needs a search, not a second opinion |
| 3. An invented source | 29 research and cite sources | A real-looking citation – author, year, title – that does not exist | No. A fabricated citation looks identical to a real one from the inside | Click every link. If it has no link, treat it as not cited |
| 4. Broken code | 26 review · 27 find and fix the bug · 28 optimize | Code that reads correctly and fails on an edge case | Somewhat – a re-read catches obvious errors | Running it. Code is the one output on this card with a free, definitive verifier |
| 5. Tone drift the low-stakes one |
3 professional · 4 more human · 20 persuasive · 21 LinkedIn | It sounds like somebody else | Not its job | Reading it aloud |
Row three is the one to sit with. A fabricated citation is indistinguishable from a real one without leaving the chat window. Asking the model to “check your answer first” cannot help, because the check happens in the same place the invention did. Command 29 is the single entry on this card that needs an action outside the conversation, and command 30 cannot be that action.
Row four is the opposite and the happiest case: code has a free, definitive verifier – you run it. That is the only one of the five where “did it work?” has an unambiguous answer, which is why code is simultaneously the riskiest-sounding category and the easiest to check.
All thirty, regrouped
The card lists them 1 to 30 in no particular order. Group them by what they actually do and eight families appear – and two of those families have exactly one or two members each, which is where the interesting entries are hiding.
| Family | The commands, as printed | Our note – read 12 October 2026 |
|---|---|---|
| Understand 6 commands |
1 Explain this like I’m 10 · 7 Explain with a real example · 9 Compare A vs B · 13 Act as an expert in [topic] · 14 Teach me step by step · 18 Give me an example | The family that needs command 30 most. Four of the six produce claims about the world. Note 13 is the only command with a bracket in it – the only one that admits it needs an input from you |
| Condense 6 commands |
2 Summarize this in 5 bullets · 6 Give me the key points · 8 Turn this into a table · 17 Turn this into notes · 19 Make this shorter · 16 Create flashcards from this | The most reliable family. Everything comes from material you supplied, so you can always check it against the source. Command 8 is the quiet winner – forcing prose into a table exposes gaps the prose was hiding |
| Rewrite 6 commands |
3 Make this professional · 4 Make this more human · 5 Fix my grammar · 20 Make this more persuasive · 21 Rewrite this for LinkedIn · 10 Find the mistakes | Command 4 is a whole genre by itself – we took apart an eight-prompt card built entirely around it and found the same gap: rewriting removes robot-tells and cannot add anything of yours |
| Study 2 commands |
15 Quiz me on this · 16 Create flashcards from this | Underrated pair. Command 15 is the only one on the whole card that tests you rather than producing something for you – which makes it the only one whose output you cannot fake your way past |
| Generate 5 commands |
12 Give me 10 better ideas · 22 Give me 10 viral hooks · 23 Give me 10 captions · 24 Turn this into a script · 25 Generate 20 content ideas | The only family that starts from nothing. Note the numbers: 10, 10, 10, 20. Volume is the promise, and volume is the easy half – picking the one good hook out of ten is still your job |
| Code 3 commands |
26 Review this code · 27 Find and fix the bug · 28 Optimize this code | Named in the footer as a risk category, and rightly. But also the only outputs here you can run, which settles the question better than any amount of re-reading |
| Meta 2 commands |
11 Improve this prompt · 29 Research and cite sources | Command 11 is the most valuable entry on the card. It is the only one whose output is another command – on a card headed “Copy · Paste”, it is the one telling you to stop pasting and start writing better instructions |
| Check 1 command |
30 Check your answer first | The odd one out, and the subject of this post. Not a peer of the other twenty-nine – a suffix for a dozen of them |
Every command is missing the same two words
Three to six words is the card’s whole design, and brevity is a real feature – you can remember these, which is more than can be said for the eighty-word prompt paragraphs that usually circulate. But look at what the brevity leaves out.
“Make this professional.” Professional to whom? A law firm and a startup mean opposite things by that word.
“Summarize this in 5 bullets.” For a decision you are about to make, or for someone who will never read the original? Those are different summaries.
“Give me 10 viral hooks.” For which platform, which audience, which video?
Every one of the thirty is missing who it is for and what it is for. Adding those two is a few seconds and it changes the output more than any other edit you can make:
“Summarize this in 5 bullets for someone deciding whether to switch editors this week, so they can make that call without reading the original.”
And this is exactly what command 11 is for. “Improve this prompt” is the only entry whose output is a better instruction rather than content – which makes it the command that fixes the other twenty-nine. On a card that says Copy · Paste · Done, number 11 is the quiet instruction to stop copying and start specifying. We made the same argument about a set of business prompts where sixteen bracketed blanks carried the whole quality of the answer.
The twelve that need command 30 welded on
Here is the practical version. Do not run command 30 as its own step – you will forget, and by then the output is already in a document. Append it, in the same message, to the commands that can hand you something false. The suffix differs by what can go wrong.
| Commands | What can be false | The suffix to add | Why this wording |
|---|---|---|---|
| 7, 13, 18 examples, expert answers |
Invented specifics stated with confidence | “…then mark anything you are not certain of, and say which parts are general knowledge rather than facts about my situation.” | Forces the answer to label its own confidence instead of delivering everything in one flat register |
| 9 compare A vs B |
Attributes of products it half-remembers | “…and list which claims about each option I should verify before relying on them.” | A comparison is a pile of small factual claims. This turns it into a checklist |
| 29 research and cite sources |
The citations themselves | “…give a working link for every source. If you cannot give a link, say so and do not name the source.” | The only fix that works. A name and a year can be generated; a link either loads or it does not |
| 26, 27, 28 code |
Logic that looks right | “…then give me the three inputs most likely to break this, and what should happen for each.” | Better than “check your answer”. It produces test cases, and a test case is checkable by running it |
| 2, 6, 17, 19 summaries and notes |
Something important left out | “…then list what you left out and why.” | Makes the compression visible. You can judge the omissions far faster than you can re-read the source |
| 10 find the mistakes |
Mistakes it did not find | “…and say what you checked for and what you did not check.” | A clean bill of health means nothing until you know what was examined |
| 12, 22, 23, 25 generate ten of something |
Nothing factual — but nine are filler | “…then rank them and tell me which one you would actually use, and why the bottom three are weak.” | Ten options is a way of avoiding a decision. Make it commit |
Row three is the one that matters most and it is the cheapest. “If you cannot give a link, say so and do not name the source.” That single clause converts command 29 from the riskiest entry on the card into one of the safest, because it moves the check out of the conversation and onto something that either loads in a browser or does not.
Row seven is the one that will improve your week. The generate commands all promise volume – ten hooks, ten captions, twenty ideas – and volume is the half that was already easy. Asking which one it would actually use costs nothing and turns a list into a recommendation.
Four honest caveats
The footer is excellent and does most of our work for us. “Always check AI answers before you use them, especially facts, sources and code. Results vary, so keep your own judgement.” It names three risk categories specifically rather than waving at caution in general, and it ends by putting the judgement back on the reader. That is the correct sentence, and our whole post is really just an argument about where on the card it should appear – not at the bottom in grey, but attached to the twelve commands it describes.
Brevity is a feature, not a flaw. Three-to-six-word commands are memorable, and a command you remember beats a perfect eighty-word prompt saved in a file you never open. We are not arguing for longer commands. We are arguing for two added words – who it is for, what it is for – and one added clause on the dozen that can be wrong.
Command 30 is genuinely useful, just mis-placed. Asking a model to re-read its own output does catch real errors – dropped caveats, arithmetic slips, internal contradictions. It is weak precisely where the stakes are highest: a fabricated citation and a real one look identical from inside the same conversation. That is not a failing of the instruction; it is a limit of self-checking as a method.
And we have not run them. Everything here is read off the card as printed. We have not measured what any particular model does with any of the thirty, and the grouping into families is our reading – command 16 sits in two of them, which is why eight families cover thirty-one slots.
The one-line version
Twenty-nine of these commands act on your material. The thirtieth acts on the answer, and it is printed last, in the same size box, as though checking were a task you do once at the end rather than a clause you attach to the things that can be wrong.
So do not run command 30. Weld it on. To the example and expert commands: mark what you are not certain of. To the citation command: a working link or no source at all. To the code commands: the three inputs most likely to break this. To the summaries: list what you left out. To the generators: rank them and pick one.
And use command 11 more than any of the others. “Improve this prompt” is the only entry whose output is a better instruction – on a card headed Copy · Paste · Done, it is the one quietly telling you that copying and pasting is the part that was never going to be enough.
Learning a lot does not make anyone good. Doing a lot does. Saving thirty commands takes one tap. Adding seven words to the one command you actually use today takes ten seconds and is the only part that changes what you publish.
If you want to be held to that, we run a free challenge at BANI Academy. There is no barrier to entry – you commit a small fee that is refunded when you finish, and the only requirement is that you actually do the work for the length of the challenge.
Run command 29 with the link clause attached, and tell us how many of the sources actually loaded. Comments are under the original reel.
More on building a channel for a global audience at mmoyoutube.com.
