The Spirit of the Instructions

Say as little as possible

The common intuition holds that a good instruction tells the model what to find. Metamorfon rests on the opposite principle: an instruction should say as little as possible about the conclusion it expects. What it prescribes is not a result but a form — axes, an order of examination, a discipline of restitution. The content is left to the debate.

The reason is simple. The wording of an instruction is not neutral: sometimes small variations in word choice or prompt structure can significantly affect an LLM’s responses (Chatterjee et al., EMNLP 2024). Naming in advance what one is looking for therefore increases the risk of making it appear in the result. Asking a model to “identify the biases” presupposes that biases are there; asking it to “map what resists” at least leaves open the possibility that nothing does.

Writing the instructions therefore consists in removing, one by one, the terms that tilt the inquiry. We call this lexical depopulation. Saint-Exupéry wrote that perfection is reached “not when there is nothing more to add, but when there is nothing left to take away.” An instruction follows the same rule: it is ready when nothing more can be removed without making it inoperative.

This caution toward lexically loaded terms matters all the more because models can also adapt their answers to preferences they infer from the user, a behavior documented under the name sycophancy (Sharma et al., 2023).

Suggest rather than command

The same restraint governs the way a dialogue instruction addresses the models. It suggests more than it commands: it proposes a working posture rather than dictating every step, and deliberately leaves some modalities implicit. This is not a loosening of the method, but a continuation of the same principle.

A highly prescriptive instruction narrows the space of interpretation; leaving some modalities open preserves more latitude. Over-specifying how the models should debate would stifle the antagonism that gives a session its value. And whatever can remain unsaid lightens the load: following many instructions simultaneously is not neutral for LLMs. Performance can depend, among other things, on the order and position of instructions (Chen et al., EMNLP 2024).

This way of writing was not decided from the outset; it was learned. Over several months of test sessions, the dialogue instructions were progressively lightened — made less imperative, more generic, stripped of directives that dictated the form of an exchange. As they receded, the debates became more fluid. The models stopped responding according to a pattern imposed from outside; relieved of part of that constraint, they focused more on the substance of the argument than on its formatting, and their voices became more distinct — each recovering a style of its own that overly tight instructions had tended to flatten. We do not claim to isolate a single cause, since the models themselves also changed over the period; but the tendency recurred regularly enough, from one version to the next, for us to rely on it.

Analysis instructions belong to a different order. Reliably reconstructing a debate does not require the same latitude as conducting one, and some of these instructions are more demanding — their function requires it. But whether an instruction suggests or holds a form, it never touches the conclusion: neither the dialogue nor the analysis receives it in advance.

This restraint in the writing of instructions is only one part of the system. How those instructions then interact with architectures, debate modes, and the steering of successive turns is developed in Make Two AIs Debate: Orchestrating and Analyzing the Dialogue.

Equity and symmetry

Lexical depopulation addresses a first level: what the instruction says. By removing terms that tilt the inquiry, it seeks to prevent a conclusion from being suggested in advance. A second, less visible level remains: what each model receives. A neutral instruction distributed unequally would cease to be neutral.

The architecture of Cross Dialogues and Cross Trilogues addresses this second level. At each turn, the debaters receive the same context and the same instructions — none starts with an advantage, none speaks from a privileged position. Lexical depopulation seeks not to steer toward a conclusion; the architecture seeks not to steer toward a model. One acts on the content of the instruction, the other on the treatment of the participants.

This has a consequence that becomes visible in the sessions. When a debate reveals an asymmetry — one model taking the initiative, setting the questions, pre-scripting its opponent while the other responds — that asymmetry becomes legible precisely because the context itself does not vary. Under symmetric treatment and identical instructions, the observed divergence is not attributable to unequal treatment; it can then be examined as arising from the behavior of the models rather than from a privilege built into the system.

The generic endures, the prescriptive ages

In On the Mode of Existence of Technical Objects (1958), Gilbert Simondon distinguished the abstract technical object, over-adapted to a particular milieu and made fragile when that milieu changes, from the concrete object, more integrated and more open. An instruction over-fitted to a particular model is an abstract object in this sense: optimized for one state of the system, obsolete when that state evolves. An instruction founded on principles rather than detailed prescriptions survives these transitions because it describes a task, not a way of solving it. We write for the models we do not yet have.

It is also a matter of independence. The system confronts competing models from different providers, by conviction as much as by method: model diversity aims, among other things, to reduce dependence on a single source. An instruction that assumes a particular model betrays that commitment. A generic instruction respects it.

Transparency

The less the instruction and the user are present in the result, the more the result belongs to the debate. Intervention exists — it can shift a discussion from one register to another — but it is a lever pulled occasionally, not a steering wheel held continuously. The value of a session lies in what the models produced among themselves, not in what was whispered to them.

From this follows a test that the system applies to itself.

When the same debate is analyzed by two independent models, they may recover the same tensions, the same positions, or the same lines of fracture while judging their depth differently. Such convergence does not prove that the instruction had no effect; it does, however, provide an indication of robustness when different analysts independently recover the same structure in the material.

The intended role of the instruction is to intervene as little as possible in the content, so that what comes from the material can be distinguished as clearly as possible from what comes from the analytical frame. It should be transparent enough for the material to show through. A successful instruction is one the reader, seeing the result, would not have guessed was there. It does not display itself in what it produces. It has done its work when it can no longer be seen.

This pursuit of transparency belongs to a broader principle: confrontation is useful only if the system does not confuse what it brings to light with what it has itself introduced. That broader question is developed in Metamorfon, an Instrument of Thought.

The exceptions

The analysis modes describe; they do not judge. Two modes are exceptions, and in opposite ways.

Argumentative Evaluation is normative by construction — but even there, the principle holds: the instruction provides criteria, not predetermined verdicts. It states what to evaluate by, never what to conclude.

Source Verification, by contrast, explicitly returns verdicts — confirmed, partially correct, incorrect, misattributed, unverified, unverifiable, or out of scope. But these verdicts are not applied to the value of ideas or the soundness of positions: they assess the degree to which factual claims are supported by sources outside the debate.

One states by what standards a line of reasoning should be judged; the other qualifies what can be established outside the debate. In both cases, the boundary between describing and judging remains explicit — and it is precisely because that boundary is explicit that it can be crossed without confusion.