← Back to blog

Self-Consistency, Meta Prompting, Least-to-Most, and PAL

Self-Consistency 🟡

Need: You need higher confidence in the answer to a complex problem.

Explanation: Generate multiple answers (via CoT) and pick the most common/best one — “voting.”

Prompt format:

“Solve the problem in 3 ways, compare the results, and choose the best one.”

QA usage example

“Propose 3 test coverage strategies for the payments module. Compare them and choose the most effective one.”


Meta Prompting 🟡

Need: You need AI to write a better prompt than you would write yourself.

Explanation: The model generates or optimizes a prompt — a prompt that creates a prompt.

Prompt format:

“Write an optimal prompt for [task] that will produce the best results with [model].”

QA usage example

“Write an optimal prompt for generating API test scenarios that accounts for edge cases, Gherkin format, and priorities.”


Least-to-Most 🟡

Need: The problem is too complex for a single prompt — it needs to be broken into smaller pieces.

Explanation: You split the problem into subproblems, solve from the simplest, and build on the results.

Prompt format:

“Break the problem into smaller parts. Solve the simplest one. Use the result to solve the next one.”

QA usage example

“How to test the entire payment flow? 1) First: form validation. 2) Then: integration with the payment gateway. 3) Finally: error handling and retry.”


PAL 🔴

Need: You need precise calculations or logic — the model assists itself with code.

Explanation: Program-Aided Language — the model generates code (Python, JS) as a reasoning tool instead of computing on its own.

Prompt format:

“Solve problem [X] using Python code. Show the code and the result.”

QA usage example

“Calculate test coverage: 47 scenarios across 12 modules, but 3 modules have 2 configuration variants each. Write Python that computes the minimum number of tests.”


In the next post: RTF, CARE, CREATE, and Other Starters — simple frameworks for a quick start.