Self-Ask | Artizan Software

The Idea

Some questions look simple but actually require connecting multiple facts. "Where was the lead singer of Foo Fighters born?" needs two pieces of knowledge: who the lead singer is, and where that person was born. AI often knows both facts individually but stumbles when it needs to chain them together.

Self-Ask solves this by having AI do what a good researcher does: recognize when a question has hidden sub-questions, ask those sub-questions explicitly, answer each one, and then combine the answers.

Building Blocks

This composition automates:

Break Down the Question Think Step by Step

Where "Break Down the Question" is a single-prompt instruction, Self-Ask turns it into an iterative loop. AI generates sub-questions one at a time, answering each before deciding what to ask next.

See It in Action

Question: "Who was president of the United States when the first iPhone was released?"

AI recognizes sub-questions are needed

AI thinks

"Are follow-up questions needed here? Yes. I need to know when the first iPhone was released before I can answer who was president then."

↓

First follow-up question

AI asks itself

"When was the first iPhone released?"

Intermediate answer

The first iPhone was released on June 29, 2007.

↓

Second follow-up question

AI asks itself

"Who was the US president in June 2007?"

Intermediate answer

George W. Bush was the US president in June 2007.

↓

Combine into final answer

Final answer

George W. Bush was president when the first iPhone was released.

Each sub-question is simple enough for AI to answer confidently. The composition — chaining answers together — is where the technique adds value.

Why This Works

Research found something surprising: as AI models get bigger, they get better at answering simple factual questions, but the gap between single-fact and multi-fact questions doesn't shrink. AI might know that the iPhone launched in 2007 and that Bush was president in 2007, but fail when asked to connect them in one step.

Self-Ask bridges this "compositionality gap" by making the connection explicit. Each sub-question is a simple lookup that AI handles well. The technique just structures the process of asking them in the right order and combining the results.

The Composition

Have AI ask itself what it needs to know first. Answer each sub-question one at a time. Use earlier answers to inform later questions. Combine everything at the end.

How to Apply This

Ask AI: "Before answering, what follow-up questions do you need to answer first?"
Have it answer each follow-up question one at a time
After each answer, ask: "Are there more follow-up questions needed?"
When all sub-questions are answered, ask for the final answer combining all the intermediate results

When to Use This

• Multi-hop factual questions: "What is the capital of the country where [person] was born?"
• When AI gives wrong answers to questions that require connecting multiple facts
• Research-style questions where you need to find one fact to know what to look up next
• When you can pair it with a search engine for each sub-question

When to Skip This

• Simple direct questions — "What's the capital of France?" doesn't need decomposition
• Math and logic problems — use step-by-step reasoning or other techniques instead
• When all the information is already in your prompt — no need to look things up if the facts are right there

How It Relates

Self-Ask is the automated, multi-turn version of Break Down the Question. Where that Level 1 technique asks AI to decompose in a single prompt, Self-Ask creates a loop: ask a sub-question, get the answer, decide if more sub-questions are needed, repeat.

It's related to Least-to-Most, which also decomposes problems but plans all the sub-problems upfront and solves them from easiest to hardest. Self-Ask is more adaptive — each follow-up question depends on what was learned from the previous one.