Spotting a bad argument is not the same skill as knowing what one is
Most people can define a straw man. Far fewer catch one being used on them, live, while they are halfway through their own sentence and thinking about what to say next.
That gap is the whole problem. Reading a list of fallacies teaches you the names. It does not teach you the thing you actually need, which is noticing — under time pressure, in a conversation you care about, when the other person is being reasonable in tone and unreasonable in structure.
So this coach does something a reading list cannot. It argues against you properly, and sometimes it argues badly on purpose.
How a debate runs
You type the motion — the proposition being argued. Anything debatable works: something from the news, something from a meeting you have coming up, something you have never been able to win an argument about at home.
You pick your side. The opponent takes the other one and says so out loud, because an opponent that hides which side it is on is doing something other than debating.
Then you go, round by round. Three, five or eight rounds in the full tool; the sample here runs two.
The part that is easy to miss
On some rounds, the opponent commits a specific, named fallacy on purpose. It is drawn from a published taxonomy of 26, in five groups — relevance, presumption, ambiguity, causal, and appeals — and it is planted at whatever difficulty you asked for. Easy plants are textbook examples. Hard ones are genuinely subtle, the sort that read as a fair point until you look twice.
You are never told which round it was. That answer is not in the page, it is not in a hidden field, and it is not in a token you could unpack. It lives on the server until one of two things happens: you call it, or the debate ends.
When you think the argument you just read was wrong in structure rather than in content, you say so — in your own words, in the same breath as your reply, the way you would in a real argument. "Hold on, that is a slippery slope, and here is my actual point." You find out immediately whether you were right.
Getting the name wrong is never scored against you. Only missing one is. And if the phrase you used is not one we have listed, we say we did not recognise it rather than marking you wrong for vocabulary we failed to anticipate.
Your own reasoning is being judged too
Quietly, and not to your face mid-argument. Every round you speak in is read against the same taxonomy the opponent draws from, and you see the verdict at the end.
That timing is deliberate. Being told you have just committed a false dilemma while you are mid-argument derails the debate and teaches worse than the same sentence does five minutes later, with the transcript in front of you.
Every charge is quoted back from your own words. If we cannot point at the sentence, the charge is dropped rather than shown — an accusation you cannot argue with is not feedback.
The scorecard
- How many plants you caught, and how many went past you. When nothing was planted it says not tested rather than 0% — you cannot fail at something nobody asked you to do.
- How your own arguments scored, round by round, at the strictness you chose.
- The fallacies you committed, each quoted from your own words, with the tell that gives it away and the move that answers it.
- Where the opponent actually conceded. It will not concede to flattery, and it will not concede to you repeating yourself louder.
- Every plant, revealed — including the ones you never spotted, so nobody leaves having quietly absorbed a bad argument.
This demo provides
- A two-round sample debate on any motion you type
- One opponent — the sceptical friend
- The full 26-fallacy taxonomy, at whatever difficulty and strictness you choose
- Calling a fallacy mid-debate, with an immediate verdict
- The complete scorecard, including every plant revealed
Nothing is kept. The sample scores you and then forgets you.
The full tool adds
- 3, 5 or 8 rounds, priced once when the debate opens — a debate that gets heated does not get more expensive
- All twelve opponents, across personal, professional, political and business families, each of which argues differently
- Sourced debates — the opponent backs its factual claims with real links, and you can put any claim up against sources
- Which fallacies keep getting past you, tracked across debates, so practice moves towards your gaps instead of repeating what you already catch
- GPT-4.1 and model selection on the advanced tier
What it will not do
It argues an assigned side. It is not stating a position of its own and it will not volunteer opinions on topics adjacent to the motion.
When it plants a fallacy it may only use one from the published taxonomy, and it may not invent statistics, studies, quotes or events to do it. A fallacy is an error in reasoning, not a false fact, and that line is written into the instructions it works from.
Planting is a visible toggle. It is on by default, because it is the thing worth practising — but it is never silent, and you can turn it off.