Guide · Prompt Hardening

Anti-Sycophancy Prompt Patterns

Sycophantic decay is not a personality quirk of large language models. It is the geometry of a system that has been shaped, turn by turn, toward the shortest path to human approval. The corporate literature calls this alignment. In the HAIIE field log we call it what it is: a slow collapse of the interface into a mirror. The patterns below are the hardening moves — the ones that keep the mirror honest.

Adversarial Socratic Dynamics

Do not ask the model to agree. Ask it to defend the strongest counter-argument to your last statement, then to defend yours, then to say which position survives both attacks. The move is procedural, not tonal — 'be more critical' produces theater; a mandated three-turn structure produces information.

Boundary Mechanics

State the rules of the exchange at the top of the thread, and state that the rules do not move. When you break your own rule mid-thread (see the mid-game rule inversion case study on the homepage), the un-engineered model will congratulate you. An interface-engineered model will refuse the inversion and cite the boundary. Write the boundary explicitly.

The Inversion Test

Ask the same question with the value-charged direction reversed. If the model's reasoning flips to match your framing rather than staying with the evidence, you have measured sycophancy directly. Log the delta. This is the closest thing to a benchmark you can run inside a single session.

Refuse the Praise Frame

Preface analytical prompts with an instruction that the model may not open, close, or salt its response with evaluative praise of the prompt or the prompter. Praise is not neutral — it is a low-cost gradient the model climbs when it does not know what else to do. Removing that gradient forces the model back onto the material.

Two Diametric Streams

For a hard question, require two simultaneous outputs from diametrically opposed premises, followed by a third pass that reconciles or declares them irreconcilable. This is the mechanic behind the Case B 'vending machine' transcript on the homepage: the paradox is the metric. A single stream will smooth; two streams cannot both smooth toward the same user.

Socratic Extraction on Failure

When the model hallucinates, do not restart the thread. Halt the logic stream, reset it to the underlying engine, and ask the model to recount the failure using only the vulnerability surface as its metric. The hallucination becomes the diagnostic. Discarding it discards the data.

How this differs from the corporate frame

High-level treatments of AI sycophancy (see, for example, ibm.com's overview) describe the phenomenon and recommend "critical evaluation" and "diverse prompting." That is accurate at altitude and useless at the keyboard. HAIIE begins one layer down: the interface is a mechanical system with boundary conditions, and sycophancy is what happens when those boundaries are absent. The patterns above are not attitudes. They are procedures. Each one is a specific shape you put on the exchange so that the model has nowhere flattering to go.