A weak argument has unstated premises, missing evidence, vague quantifiers, and unanswered counter-arguments, and it usually goes uncaught until a hostile reader gets to it. These 12 prompts walk you through each weakness in turn so the final version holds up to a skeptical reader. Pair with essay thesis prompts if the thesis itself is still drifting, and with counterargument generation prompts once you want to attack the draft from the outside.
TL;DR
- Run prompts 1-3 first (evidence, premises, logic gaps) to map where the argument actually breaks, then 4-7 to harden it against opposition.
- Paste your real draft into the
{paste argument}slot every time. These prompts critique your text; generic responses mean you gave it generic input. - Use a reasoning-strong model: as of June 2026 that means Claude Opus 4.7 or GPT-5.5 in Thinking mode for short essays, and Gemini 3.1 Pro or Opus 4.7 (both 1M-token context) when the argument spans a long document.
- Don’t accept the model’s verdict blind. It will flag plausible-sounding leaps that aren’t real and miss some that are. Treat output as a checklist, not a grader.
Which model to run these on
Detecting logical fallacies and unstated premises is genuinely hard for language models. Research on fallacy detection (see the ACL/arXiv work on counterargument-aware prompting) found that models do meaningfully better when the prompt asks them to construct the counterargument and explanation first, which is exactly how prompts 3, 4, and 7 below are built.
As of June 2026, for the reasoning these prompts demand:
| Model | Strength here | Context | Notes |
|---|---|---|---|
| Claude Opus 4.7 | Deepest critique of premises and concessions | 1M tokens | Best at the steelman and concession passes |
| GPT-5.5 (Thinking) | Strong logical-leap and causal checks | ~320 pages in-app on Plus | Pick the Thinking mode in the picker, not Instant |
| Gemini 3.1 Pro | Long documents at lower API cost | 1M tokens | Good when you paste a whole chapter or report |
Plus tier is $20/mo (ChatGPT), Claude Pro is $20/mo, Google AI Pro is $19.99/mo (the tier formerly sold as “Gemini Advanced”). All three handle these prompts on a paid plan; free tiers will hit message caps fast on a long argument. For a fuller breakdown see ChatGPT vs Claude vs Gemini.
Best for
- Long-form essays and op-eds
- Memos at work where decisions ride on the argument
- Debate prep and structured disagreement
- Founder pitches and product reasoning
- Grant or research proposals being read by adversarial reviewers
1. Evidence-gap audit
Below is my argument. For each claim, mark:
(a) is there cited evidence
(b) is the evidence relevant to the claim, not adjacent
(c) what stronger evidence would change the verdict
Flag any claim supported only by intuition, anecdote, or "everyone knows".
{paste argument}2. Unstated-premise finder
Below is my argument. List every unstated premise it depends on. For each premise:
- One line stating it explicitly
- One line on how a hostile skeptic would attack it
- One line on whether I should state it openly or shore it up silently
{paste}3. Logical-leap detector
Below is my argument. Find every place where the logic jumps without sufficient justification. For each leap:
- Quote the 2 sentences that bracket it
- Name the missing intermediate step
- Suggest the smallest addition that closes the gap
{paste}4. Strongest-counterargument generator
Below is my argument. Write the strongest counter-argument a smart skeptic in {field} would mount. ~200 words, no strawmen, cite the kind of evidence they would actually use.
Then write my best response in 150 words — concede what is true, push back where the response holds.
{paste}5. Steelman pass
The position I disagree with: "{opposing view}".
Write the strongest steelman version in ~200 words — better than its actual proponents usually argue it. Then mark which parts of my current argument actually engage the steelman vs which parts only engage the weak version.
{paste my argument}6. Vague-language replacer
Below is my argument. Highlight every vague word ("often", "many", "some", "typically", "tend to") and either:
- Replace with a specific quantifier or example
- Or, if specifics are not knowable, rewrite to acknowledge the uncertainty explicitly
Output a redlined version with the changes marked.
{paste}7. Concession-strengthener
Below is my argument. Find the strongest point of opposition and add 1 genuine concession there — one that would actually weaken my case if left unanswered.
Then add a response that re-establishes the conclusion. The concession must be real, not a fake "of course critics would say X" setup.
{paste}8. Causal-vs-correlational check
Below is my argument involving cause / effect. For each causal claim:
(a) is it causal or just correlational
(b) what alternative explanations exist (reverse causation, confounder, selection effect)
(c) what evidence would distinguish them
(d) is the evidence I cite that kind of evidence
{paste}9. Generalizability check
My argument: "{thesis}". The evidence I cite comes from {context — geography, time period, industry, sample}.
Evaluate: does this evidence generalize beyond that context? Name 2 specific conditions under which it would not, and what readers in those conditions should take away instead.10. Quantification pass
Below is my argument. Find every claim that should be quantified but is currently qualitative. For each:
- Suggest the right metric or proxy
- State the level of certainty I can honestly claim
- Rewrite the sentence at that certainty level
{paste}11. Stakes-and-falsifiability check
My argument: "{thesis}".
What evidence would change my mind? Be specific — a study result, a market outcome, a counter-example.
If the honest answer is "nothing", my argument is unfalsifiable and I should rewrite the thesis to be narrower or more conditional. Help me identify falsification conditions I would actually accept.12. Conclusion-sharpener
My argument's current conclusion: "{conclusion}".
Rewrite 3 stronger versions:
(a) more specific — name the population, time horizon, magnitude
(b) more honest about scope — flag what it does NOT claim
(c) more action-oriented — what a reader should do tomorrow
Then mark which fits the rest of the argument best, and which I should adopt if I want the post to be quotable.Common mistakes
- Citing evidence that is adjacent but not relevant to the actual claim.
- Strawmanning the opposition instead of engaging its strongest version.
- Inferring causation from correlation without naming alternative explanations.
- Using vague quantifiers (“many”, “often”) that hide the absence of real numbers.
- No falsification condition: the argument is structurally unfalsifiable, and therefore unverifiable.
- Trusting the model’s pass as a verdict. A fallacy detector that says “no issues” is often wrong; the value is in the specific leaps it names, which you then check yourself.
FAQ
Which prompt should I run first? Start with the evidence-gap audit (1), then the unstated-premise finder (2) and logical-leap detector (3). Those three show you where the argument actually breaks. Save the steelman (5) and concession (7) passes for after you’ve patched the obvious holes, or they’ll just pile up work you’d have redone anyway.
Will the model invent fallacies that aren’t there? Yes, regularly. Language models still over-flag plausible-sounding “leaps” and miss real ones, so the output is a checklist of places to look, not a grade. Read each flagged item and decide for yourself whether the gap is real before you rewrite.
Do I need a paid plan? For a short essay, free Claude (Sonnet 4.6) or free ChatGPT (GPT-5.5 with tight limits) will get through a couple of these prompts before hitting a cap. For a long argument you paste in full, a $20/mo plan with 1M-token context (Claude Pro, or Gemini 3.1 Pro via Google AI Pro at $19.99/mo) is worth it, as of June 2026.
Can I run all 12 in one chat?
Run them in sequence in the same thread so the model keeps your argument in context, but feed each its own fresh {paste argument} after a rewrite. Batching all 12 into one prompt produces shallow, generic output.
Why does it keep giving me generic feedback?
Almost always because the {paste argument} slot wasn’t filled with your actual text, or the draft is so abstract there’s nothing concrete to critique. Paste the real paragraphs, and name the {field} so the model knows what a skeptic in that area would actually say.