Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

adversarial-prompts

Prompts that make AI argue against you — not agree with you.

For investment theses, business plans, essays, and system designs where you want a hostile second opinion instead of a cheerleader.

Why this exists

LLMs default to agreement. This is fine for brainstorming, dangerous when you want to stress-test a decision. These prompts replace the default with a structured 5-section rebuttal that completes whether or not the easy critiques are obvious — so the hard ones get said.

The four prompts

Each prompt assigns a hostile persona, demands a fixed five-section structure, and refuses to truncate before the last section.

Domain Hostile persona The five sections
investment Senior partner at a top-tier hedge fund Weakness of the narrative / The opposing story / Missed risks / Conflict with the news / Three questions to re-ask yourself
business Partner at a venture firm writing the pass memo Weakness of the hypothesis / The opposing business case / Missed market risks / Conflict with competitor moves / Three falsification tests
essay Hostile peer reviewer with domain knowledge Weakness of the reasoning / The opposing argument / Missed counter-evidence / Problems with the cited sources / Three questions to strengthen the essay
code design Senior staff engineer assigned to play devil's advocate Weakness of the design / The opposing architecture / Missed failure modes / Conflict with established best practice / Three review checkpoints to clear before merging

Quickstart

  1. Open the prompt for your domain in prompts/.
  2. Substitute your values into the {{placeholder}} slots inside the prompt body.
  3. Send the filled prompt to a capable LLM.
  4. Read the rebuttal slowly, in order. The first section names what is fragile; the last section names what to monitor. The middle three exist so the first and last hold up under scrutiny.

A minimal example for the investment prompt:

# (paste the entire prompt body from prompts/investment-thesis.md here)

# Input

Company: NVDA (NVIDIA Corporation)
Thesis: NVIDIA's lead in AI infrastructure is structural ... (your thesis)
Evidence: Q3 2025 data center revenue +94% YoY ... (your evidence)
Exit Condition: Sell if any hyperscaler ... (your exit condition)
Recent News:
- Amazon expands Trainium 2 production ...
- Google announces TPU v6 ...

No CLI, no SDK, no install step. The whole repo is plain markdown and copy/paste.

Examples

Each domain ships a worked example pair: a deliberately imperfect input plus the verbatim model rebuttal. See examples/README.md for the reading guide.

Tested across Claude Sonnet 4.6 and Claude Opus 4.7

The investment example was generated by Claude Sonnet 4.6. The business, essay, and code-design examples were generated by Claude Opus 4.7. Both models produced the full five-section structure without modification when the prompt was used as written.

Each example file carries a YAML front-matter block declaring which model produced it, with the prompt path, the generation date, and a note on verification. Independent verification with Gemini, GPT-5, and Llama-class models is welcomed — open a PR and we'll add your reproduction to the repo.

Recommended model settings

  • max_tokens: 4096 or higher. The five-section structure runs long; lower ceilings produce silent truncation.
  • Temperature: 0.7 - 1.0. Lower temperatures collapse into boilerplate ("the user should consider risks") instead of producing specific critique.

Design philosophy in one paragraph

The load-bearing trick is the fill-in-the-blank output template embedded in every prompt. Without it, models drift into polite summaries. With it, the model copies the template into its output and fills each slot, which forces completion of all five sections instead of trailing off after the easy critiques. Two reinforcing rules make the template hold: a required citation of a historical failure case in every rebuttal (so the critique grounds in something concrete instead of generic "have you considered" hand-waving), and an explicit rule that the model must finish through the last section (which catches the early-termination failure mode that long-context responses otherwise hit).

For a full section-by-section walkthrough, see docs/prompt-anatomy.md. For the case against sycophancy and the alignment-research lineage these prompts inherit from, see docs/why-adversarial.md.

Limitations

  • Models occasionally fabricate historical failure cases. Verify any cited precedent before quoting it externally.
  • Single-model adversarial critique cannot substitute for independent perspectives; treat the rebuttal as a sparring partner, not an oracle.
  • These prompts test your thinking. They do not replace it.

Contributing

New domains are welcome. The seven blocks that every prompt in this repo shares are documented in docs/prompt-anatomy.md — start from one of the existing prompts and re-use the skeleton.

Pull requests that report results from running these prompts on Gemini, GPT-5, Llama-class, or Mistral-class models will be merged with credit. Pull requests that add new example pairs (input + rebuttal in the same shape as the existing four) are also welcome.

Issue reports about prompt regressions on specific models are particularly valued. If a model starts producing only four of the five sections, or starts adding emoji headers, or starts hedging into "this is not advice" disclaimers, that is a fixable prompt drift.

Acknowledgments

These prompts inherit from three older ideas.

  • Anthropic's Constitutional AI research showed that a model can be trained to critique its own outputs against a written constitution. The template-driven adversarial role here is a downstream descendant of that intuition — applied at prompt time instead of at training time.
  • OpenAI's "AI safety via debate" paper (Irving, Christiano, Amodei, 2018) framed productive disagreement between aligned agents as a path to better answers than monologue produces. These prompts are a single-model echo of that pattern.
  • Cunningham's Law ("the best way to get the right answer on the internet is not to ask a question, it's to post the wrong answer") generalizes: the fastest way to find what is wrong with your thesis is to make the AI argue the opposite as if it believed it.

License

MIT. Fork, remix, ship. If you build something on top, opening an issue linking to it would make the maintainer's day.

Maintainer: @inunyokki-cmd

About

Prompts that make AI argue against you — not agree with you. Investment theses, business plans, essays, and system designs.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors