A four-step AI framework for preparing high-stakes conversations

Executive overview

Most people rehearse tough conversations in their head — then get thrown off the moment the other person says something unexpected. AI defaults to agreeing with you, which makes the problem worse.

SPAR turns any AI into a sparring partner: it clarifies your position, builds the strongest case against you, and lets you practice the real conversation before you're in the room.

The core insight: walking in having already had the conversation five times changes everything.

Sharpen — get clear on your own position

• Use an AI interview: the AI asks one question at a time, each informed by your previous answer.
• Use dictation, not typing — ranting gives the AI far more context and nuance.
• Cap the interview at 20–25 questions (fewer leaves gaps; more causes fatigue).
• Prime the AI to clarify three things: ideal outcome, acceptable compromise, non-negotiable.
• At the end of the interview, switch to a high-reasoning model to synthesise a structured summary — primary objectives, secondary objectives, and non-negotiables.

Probe — understand the other side better than they do

• Paste your structured summary into a fresh conversation with a high-reasoning AI.
• Ask it to build the strongest possible case against your position from a reasonable, intelligent person's perspective.
• Follow up in the same thread: ask for the five most likely ways the other person could react to your points, with specific responses for each scenario.
• Skipping this step is the most common prep mistake — AI makes it tractable.

Attack — find every hole in your argument

• Use a separate AI conversation; frame it as adversarial critique, not coaching.
• Three intensity levels:
  1. Devil's advocate — challenge the plan, point out flaws and weak assumptions.
  2. Red teamer — actively hunt for weaknesses, expose blind spots.
  3. Extreme skeptic — tear apart the argument, no softening, feedback must be specific and actionable.
• Run the attack through two or three different models; each surfaces different types of holes.
• Model selection matters: Claude tends to be the softest; Grok the most intense out of the box.

Rehearse — have the conversation in voice mode

• Use outputs from all three previous phases to create an AI persona via a system prompt.
• Ask an AI to research best practices for prompting your target model, then generate a system prompt that embeds the persona and a tiered intensity scale (e.g. level 3 = friendly pushback, level 6 = pointed questions, level 9 = hostile).
• Embed the system prompt in a ChatGPT project, Gemini Gem, or Grok project, then open it in voice mode.
• Set a trigger word (e.g. "purple rhino") so the AI only interrupts when you're actually done speaking.
• Run three to six practice conversations; use the best one to build a battle card.

Building the battle card

• Paste your best conversation transcript plus the probe-phase reaction scenarios into a new prompt.
• Ask the AI to produce a one-page cheat sheet with:
  • Strongest opening line
  • Three key phrases that landed well
  • Bridging statements for when the conversation goes off-track
  • Pre-drafted responses to the most likely attacks on your position