A good prompt is a small argument plus a request, and so is any clear ask in life. Give your premises, terms, limits and goal, and there is less to guess. It can still be wrong, so check.
A chatbot has never met you. Everything it knows about your request is what you wrote. These are the ideas logicians use to make a line of thought easy to follow, and each one has a job in a prompt.
A statement you take as given and reason from. If a premise is false, anything built on it can be false too.
'The museum is closed on Mondays' is a premise. Your plan for Monday starts from it.
In a prompt: Your context, examples and definitions are the premises. The model reasons from what you give it, or from what it guesses.
An inference is the step from premises to a new claim. The conclusion is where that step lands: the thing you wanted to show or decide.
The museum is closed on Mondays. Today is Monday. So the museum is closed today.
In a prompt: Your objective names the conclusion you want. The answer is the model's step from your premises toward that goal.
The set of things a statement is talking about. Words like 'all' and 'everyone' only make sense inside a chosen domain.
'Everyone passed the test' is about one class, not everyone on Earth.
In a prompt: Scope sets the domain. Without it, the model picks one for you, often wider or narrower than you meant.
A definition fixes what a word means. A word is ambiguous when it could mean more than one thing. Sliding between meanings mid-argument is called equivocation.
'A light meal' could mean low in calories or just small. Saying which one removes the doubt.
In a prompt: Terminology defines your key words, so 'cheap', 'short' or 'simple' mean what you mean.
A necessary condition must be met, but may not be enough. A sufficient condition is enough on its own.
A ticket is necessary to ride the train, but not sufficient: the train also has to be running.
In a prompt: Limits, format and time frame are necessary conditions. Meeting them all still does not guarantee a good answer.
Evidence is what supports a claim. Induction draws a general pattern from particular cases. It can be strong, but it is never certain.
Every swan you have seen is white, so you expect the next one to be. Then you visit Australia and meet a black swan.
In a prompt: Examples teach by induction. Citations, statistics and quotes are evidence, and they only count if they are real.
A single case that shows a general claim is false. Hunting for one is how you test an idea instead of just believing it.
'All birds can fly.' Penguins are a counterexample, and one is enough.
In a prompt: Counterarguments and other points of view ask the model to look for the case where your idea fails.
Philosopher H. P. Grice described what cooperative talk assumes: say enough but not too much, say what is true, stay relevant, and be clear.
'I'm out of petrol (gas).' 'There's a garage (a gas station) round the corner.' You assume it is open and sells fuel. That is cooperative talk at work.
In a prompt: A good prompt gives the model what it needs to be informative, truthful, relevant and clear for your purpose.
Grice described talk between people. Using his maxims on a chatbot is a useful lens, not a claim that the model is a cooperative person.
Both arguments have the same valid form: if the premises were true, the conclusion would have to be. Only the first is sound, because its premises really are true.
The same goes for prompts. A tidy, well-structured request can still be built on a wrong premise, and the answer will follow it faithfully.
Twenty-one parts, in seven logical families. You never need all of them at once. Each one you leave out is something the model has to guess. Tap any element to read it.
A role for the model to take on, which sets its vocabulary, depth and point of view.
The logicFixes the speaker's standpoint. Like choosing a perspective in an argument, it shapes which premises seem relevant and which words get used. It adds no knowledge or credentials.
Leave it out and the model guesses Which expert voice, level of detail and angle you want.
Watch out: A role is a costume, not a qualification. A model playing a doctor is not a doctor, so check health, legal or money advice with a real one.
Who will read or use the answer, and what they already know.
The logicSets the shared background the answer can assume. What counts as informative enough depends on the listener, so the audience fixes how much to explain.
Leave it out and the model guesses Whether to explain the basics or skip them, and which words the reader knows.
Watch out: A vague audience like 'beginners' helps less than saying what they already know and what they need to do with the answer.
The emotional colour of the reply: friendly, formal, firm, playful or calm.
The logicTone changes how something is said, not what is claimed. Logic separates content from expression, and tone is a request about expression.
Leave it out and the model guesses Whether to sound warm, neutral, formal or blunt.
Watch out: Tone can blur content. If you ask for 'reassuring', check the reply has not softened a fact you needed to hear plainly.
Which language to answer in, and how plain or technical the wording should be.
The logicA statement only works if both sides read it the same way. Choosing the language and reading level avoids obscurity, which Grice's maxim of Manner warns against.
Leave it out and the model guesses Which language, spelling and reading level to use.
Watch out: A translation can sound fluent and still be wrong. For anything important, like medicine labels or contracts, ask a fluent speaker to check it.
The background facts of your situation that the answer should build on.
The logicContext supplies the premises: the statements taken as given. A conclusion is only as good as the premises it starts from.
Leave it out and the model guesses Your situation, so it fills the gap with a typical case that may not be yours.
Watch out: False premises give confident nonsense. If your background facts are wrong or out of date, a well-reasoned answer is likely to be wrong too.
One or more samples of what a good answer looks like.
The logicExamples work by induction: from a few cases the model infers a general pattern. Like any induction, the pattern it picks may not be the one you meant.
Leave it out and the model guesses The style, length and pattern you have in your head.
Watch out: Models can copy examples closely, including their accidents. If all your examples are food words, expect more food words. Vary them on purpose.
Words or names that must appear, or that pin down the topic.
The logicKeywords mark the subject terms the answer must be about. They work as a relevance test: an answer that never touches them has drifted off the question.
Leave it out and the model guesses Which specific names, things or ideas you care about.
Watch out: A long keyword list turns into word salad. Pick the few terms that truly matter, and check they are used correctly, not just dropped in.
Definitions for the words in your request that could mean more than one thing.
The logicDefining terms blocks equivocation, the error of letting one word shift meaning partway through. A definition fixes what a key word refers to from start to finish.
Leave it out and the model guesses Which meaning of an ambiguous word you have in mind.
Watch out: Your own words feel obvious to you. Words like 'cheap', 'soon', 'healthy' and 'short' are where misunderstandings hide. Pin them down with numbers.
What the answer should cover, and what it should leave out.
The logicScope sets the domain of discourse: the set of things the conversation is about. Claims are judged inside that domain, so a fuzzy domain gives fuzzy answers.
Leave it out and the model guesses How wide or narrow to go, and what counts as off topic.
Watch out: A scope that is too narrow can hide what matters. Invite the model to flag anything important just outside your boundary.
Rules the answer must obey, such as length, cost or things to avoid.
The logicLimitations are necessary conditions: an answer that breaks one fails, however good it is otherwise. Meeting every limitation still does not make an answer good enough.
Leave it out and the model guesses Your hard limits, so it may suggest things you cannot use.
Watch out: Check the answer against each limit yourself. Models sometimes quietly break a rule, such as a word count, while sounding as if they kept it.
The time frame your task, or the information you need, belongs to.
The logicMostly practical, not logical. A chat model answers when asked, not on a schedule. The logical part is a time index: many claims are true only as of a date.
Leave it out and the model guesses Which dates matter, and how current the facts must be.
Watch out: Models have a knowledge cutoff and may not know today's date. For anything time-sensitive, give the date yourself and check recent facts elsewhere.
The shape of the answer: a list, a table, steps, an email or a short paragraph.
The logicFormat is a constraint on form, not content. It answers Grice's advice to be orderly: the same information is easier to check in a shape you can scan.
Leave it out and the model guesses Whether you want a list, a table, a paragraph or steps.
Watch out: A neat table can make guesses look like facts. Format makes an answer easier to read, not more reliable.
What you want the answer to achieve: the point of the whole request.
The logicThe objective is the conclusion you are reasoning toward. Without it, relevance has no anchor, since Grice tied relevance to the purpose of the conversation.
Leave it out and the model guesses Why you are asking, so it cannot tell what is useful.
Watch out: Stating a goal can tilt the answer toward it. Ask 'help me decide' rather than 'convince me', unless you really want only one side.
The next step the answer should end on, for you or for the people who read it.
The logicAn argument can end in a practical conclusion: a decision about what to do. A call to action turns the reasoning into one clear next step.
Leave it out and the model guesses Whether to end with a next step, and what that step is.
Watch out: Pushy calls to action backfire. Ask for one clear, honest next step, not pressure or invented urgency like 'only two places left'.
Sources the answer should point to, so you can check where each claim comes from.
The logicA citation links a claim to its evidence. It serves Grice's maxim of Quality: do not say things you lack adequate evidence for.
Leave it out and the model guesses Whether you need sources, and which kinds you trust.
Watch out: Models can invent citations that look real: plausible titles, authors and links that do not exist. Give your own sources when you can, and open every citation before trusting it.
Numbers that support a claim, such as counts, averages or rates.
The logicUsing a number to support a general claim is a form of induction. A number supports a claim only if it is real, recent and about the right group.
Leave it out and the model guesses Whether you want numbers, and how precise they must be.
Watch out: Models can produce precise-looking numbers with no real source. Supply your own data where possible, and trace every number to its origin before you repeat it.
Exact words from a real person or text, repeated word for word.
The logicA quotation is testimony: evidence that someone said something. It counts only if it is accurate and in context. A misquote is a false premise.
Leave it out and the model guesses Whether you want quotes, from whom, and how exact.
Watch out: Models can invent quotations or pin real ones on the wrong person. Paste the exact words yourself, or check any quote against a reliable original.
Other people's perspectives the answer should consider, not only yours.
The logicLooking from several standpoints tests whether a conclusion holds for everyone affected. A premise that seems obvious from one side may be false or irrelevant from another.
Leave it out and the model guesses Whose interests and views count, so it may give only one side.
Watch out: Listing views is not the same as weighing them. Two sides shown as equal can hide that one has much better evidence.
The strongest objections to an idea, so you can test it before relying on it.
The logicPopper's idea of falsification: try to refute an idea, not only confirm it. One counterexample can sink a general claim, and objections look for the weak spots.
Leave it out and the model guesses Whether you want honest pushback or simple agreement.
Watch out: Models tend to agree with the person asking. If you never ask for objections, the lack of any is not evidence your idea is good.
A comparison with something familiar that helps explain something new.
The logicAn analogy maps the structure of a known case onto a new one. It explains and can suggest, but it never proves; test it by finding where it breaks.
Leave it out and the model guesses Which familiar things you know well enough to compare against.
Watch out: Every analogy breaks somewhere. Ask the model to say where its comparison stops working, so you do not carry the wrong part over.
An invitation for the model to ask you what it needs before it answers.
The logicQuestions turn a one-shot guess into a dialogue. Missing premises get supplied instead of assumed, so the answer can be as informative as you need.
Leave it out and the model guesses Everything you forgot to say, silently filled in with typical defaults.
Watch out: The model's questions are only as useful as your answers. Answer honestly and specifically; if you guess, it will build on your guess.
More parts is not always better: “everything” makes a long prompt. The skill is noticing which guesses matter for this task and closing those.
Each request below got a disappointing answer. Which single part, added to the prompt, would have helped most?
A well-built prompt makes a useful answer more likely. It does not make the answer true. A valid argument from false premises can reach a false conclusion with nothing in its form to warn you, and a model can reason smoothly from a wrong guess.
When asked for citations, quotes or statistics, models sometimes produce ones that look real but do not exist. Give your own sources and data where you can, and check every reference, quote and number before you rely on it or share it.
A model learned from text up to a certain date. It may not know what happened since, or even what today's date is. Say 'as of' when time matters, and check recent facts with a current source.
Telling a model to act as a doctor, lawyer or accountant changes its vocabulary and angle. It does not give it training, a licence, accountability, or knowledge of your full situation.
The answer is a draft for you to judge, not a decision made for you. For health, legal, money or safety questions, check with a qualified person before you act.
Before you press send, run down the families. Ticks stay in this browser only.