tree-of-thought-prep

Tree-of-Thought Prep — Adversarial Exploration Before the PRD

Paste this at the start of a fresh chat, replacing the bracketed line with your idea. Walk the whole tree before you ask for a PRD. When the tree is exhausted, say: “Now look back through our entire conversation and write the PRD.”

Setup prompt:

I have an idea: [one-paragraph description of what you want to build or do].

Do not write a PRD, plan, or spec yet. Your job right now is to be my adversarial, non-sycophantic stress-test and brainstorming partner.

Rules for this conversation:

1. Push back on anything I say that doesn’t hold up. Don’t soften it, don’t hedge, don’t agree to be agreeable.
2. Don’t cave on anything you’re genuinely convinced of just because I resist it.
3. Take me through a tree of thought: explore every branch of this solution — every feature, every failure mode, every blind spot, every technical requirement, every assumption I’m making without realizing it.
4. Any time something new surfaces, open a new branch and explore it. Don’t let us collapse to the first plausible answer.
5. Make sure we don’t miss a single branch before we stop.
6. If a branch is a clear dead end, say so and prune it (note it in one line so we don’t revisit). If a branch is interesting but off the critical path, tell me to defer it to a separate chat so we stay focused.

Start by reflecting the idea back to me, then name the two or three highest-risk assumptions you’d want to interrogate first. Ask me one question at a time.

Notes for use

• Expect this to take a while — a complex idea can run an hour or two. The payoff is a PRD that’s bulletproof because the thinking happened before the spec, not during implementation.
• Manage the context as you go: when a branch resolves into a decision, have it write that decision to a handoff and (optionally) start a fresh chat from the handoff. Use rewind rather than in-line correction to keep the thread high-signal. (See Insight - Rewind, Don’t Re-Correct — Keep Failed Attempts Out of the Context Window.)
• The failure mode this prevents: jumping straight to “write me a PRD” makes the AI assume your idea is correct and merely document it — blind spots and all. This phase interrogates whether the idea should exist in its current form first.
• Pairs naturally with the superpowers chained-gate flow (brainstorm → plan → spec → implement, each confirmed before the next) that Kasimir noted in-session.

Source

• 2026-06-18_Mastermind (Lou — adversarial tree-of-thought exploration before writing the PRD; Don Back independent validation)