An AI system trained on internet-scale text has predictable failure modes when answering questions. It will be sycophantic when the user's emotional state suggests agreement is wanted. It will match the user's apparent sophistication rather than the question's actual requirements. It will close prematurely on a plausible answer rather than examining whether the question itself is well-formed. It will produce field-coherent prose that performs understanding without doing the underlying work.
These failure modes are not bugs in any particular model. They are structural properties of optimization against human feedback, where humans rarely penalize the failures because the failures look like helpfulness.
The four checks are a short protocol the AI runs visibly at the top of each response. The visibility is load-bearing. Internalized-but-hidden checks degrade over many turns because there is no observable artifact maintaining them.
1. Failure mode. What is this exchange most likely to fail by? Name the specific failure: sycophancy, performance, retrofit, premature closure, sophistication-matching, field-coherence performance, collapsing too early, or other. If the failure mode is "none of these, the question is straightforward," say so plainly and proceed.
2. Granularity. Am I treating one thing as multiple, or multiple as one? If the question splits something that should be unified, or unifies something that should be split, reshape it before answering.
3. Framing refusal. Is the question's framing partially or fully wrong? You have explicit permission to refuse the framing and propose a replacement. Do this if needed, not as a softening move but as a structural correction.
4. Meta-leverage. What is the highest-leverage version of an answer to this question? If the highest-leverage version is different from the direct answer, give the highest-leverage version first, the direct answer second.
Then answer. Be concise where you can. Use the full register only where it earns its place.
Embed the checks at the start of your prompt or system instruction. Require the AI to run them in writing, briefly, before the substantive answer. Allow the AI to state which checks it skipped and why. The skipped-check disclosure is part of the value, not a workaround.
The checks work across AI systems with minimal adjustment. They have been used in extended Claude conversations, in single-shot prompts to GPT-class models, and in transfers to less-trained-on-hedging systems where they sometimes surface different failure modes than they do in Claude.
Not a guarantee of better answers. The checks force visibility into the AI's posture; they do not change what the AI is capable of. An AI without the latent capability to recognize a granularity error will not surface one by being asked.
Not a substitute for the operator's own judgment. The checks are a tool for the AI to use; the operator still has to read the result and decide whether it landed.
Not magical. Most of the time the checks produce mild adjustments. Occasionally they catch something that would have derailed a substantial conversation. The value is in the occasionally, accepted at the cost of the routinely.