Prompt templates with real prompt engineering.

Make the model show its work before answering — dramatic accuracy gain on multi-step problems.

Explore 3 branches per decision, prune ruthlessly, beat one-shot answers.

The single move that makes any output 2× better. Used in agentic systems.

Simulate 5 experts arguing — get a richer answer than any one expert could produce.

Imagine the project failed 6 months from now. Work backwards. (Used at Amazon.)

Break the problem down to atomic facts. Rebuild from atoms. Catch the assumption everyone missed.

Run your decision through 8 named biases. Catch your own blindspot before you commit.

Generate the few-shot examples that make any prompt 2× more reliable.

Plan the full solution first, then execute. Beats one-shot CoT on multi-step problems.

Derive the general principle first. Then apply it. Fewer hallucinations, better generalization.

Reason the same question 5 different ways. Take the majority answer. Beats single-sample CoT.

One model generates, another (or the same in a different role) verifies. Catches plausible-but-wrong output.

Don't ask 'which is best?'. Ask all N×(N−1)/2 pairwise comparisons. Far more reliable.

Generate the outline first. Then expand each section in parallel. Fast and consistent on long generations.

Argue the OPPOSITE first. Then state your position with the strongest counter-argument already disarmed.

Iteratively ask 'why' 5 times. Dig past symptoms to the real cause. (Toyota Production System.)

Force separate perspectives — facts, feelings, risks, benefits, creativity, process. (Edward de Bono.)

Reframe a new problem as an instance of an old solved one. Unlocks pattern transfer.

Design a production-grade system prompt with role, rules, refusals, and meta-instructions.

Thought → Action → Observation loop. The standard pattern for tool-using agents.

Define principles up front, generate the answer, then audit it against each principle. Anthropic-style safety.

Show the AI exactly what NOT to do. Often more useful than positive examples.

Compare options without the model just agreeing with you.

Pressure-test an idea before you build it.