How does the model follow instructions, and what is the core limitation of this process?	It follows instructions via Next Token Prediction (pattern continuation), but this means there is a gap between the user's intent and the literal output.
What is the "capability zone" for steerability, where control is tightest?	Short, concrete, verifiable instructions, such as format specifications or length limits.
What type of tasks fall into the "limitation zone" of steerability?	Long reasoning chains, abstract or ambiguous instructions, or anything requiring native numerical/logical precision.
What is "reasoning drift" in the context of steerability failure?	When small, initial errors compound over multiple dependent steps in a reasoning chain.
Define "letter-over-spirit" failure.	When an instruction is followed literally, but the resulting output fails to meet the user's actual intent or goal.
What is the most effective prompting strategy for complex tasks?	Prompting from the explicit *goal* (what you are trying to accomplish) rather than just the desired *format* (what the output should look like).
Name four product features designed to improve steerability and maintain user intent.	System prompts, code execution, visible reasoning, and structured output modes.
If an instruction is followed literally but uselessly, what should the user do?	Restate the goal explicitly alongside the instruction.
What is the purpose of inserting a "checkpoint" in a multi-step task?	To verify the output or result at a critical intermediate step before allowing the model to continue.
What is the primary difference between a "goal" and an "instruction" when prompting?	The goal is the high-level objective (e.g., "Convince my team this timeline is realistic"), while the instruction is the specific format or action (e.g., "Three bullet points").