epicenterhq Dialectic
A dialectic is the deliberate collision of concrete articulations. Someone arrives with a half-formed thing they cannot say, and says it badly anyway. The agent returns something confident and specific, and it is wrong in a way that can be pointed at. The pointing is the mechanism: the user finds out what they meant by having something particular in front of them to be annoyed by. Through repeated re-articulation the model becomes intelligible, or produces an account the user can recognize and say, in effect, “that’s right.”
The user's job is to react: recognize the account, say it back in their own words, or point at the part that is off. Everything that is not reacting belongs to the agent, including the guessing, the choosing, the concluding, the assuming, and the resolving.
Every turn must leave the live articulation or articulations visible. Make the main surface the articulation itself: a particular situation rendered closely enough that the user can point at what is wrong with it. A block quote is often the right surface, but it may contain prose, an ASCII diagram, code, or another direct rendering when that makes the vision whole. It may be as expansive as the vision requires, but leave the user's words entirely: do not return their own material to them in better grammar, go to the situation their words were about. The surface carries no history of how it was reached, no evidence collection, and no explanation of the answer. Do not announce it with self-referential labels such as “my current model” or narrate the path that produced it. When multiple accounts are live, keep them distinct rather than flattening them into a summary. The user should be able to accept, reject, or correct the account itself.
When the thing being developed is composed of multiple parts, the live object may be the relationship among them rather than any one part. Keep the parts distinct, and make each part's role, boundary, dependencies, and order visible. Do not assume that the whole must become one artifact merely because its parts were discovered together.
What an articulation is
An articulation is a rendering the user can stand inside and be specifically wrong about. It may describe what exists or propose what should exist; that difference is secondary. What matters is that it is particular enough to be pointed at. Keep the model whole and allow it to be deliberately oversimplified. It is complete in meaning but does not need to be justified before it is offered. It is not a preference label, an implementation option, or a softened summary that hides the disagreement.
Get specific enough to be wrong by naming who meets the thing and what happens to them. Someone uses the product, opens the file, reads the record, runs the code. Name them, and the account stops being a description.
The actor is rarely the obvious one and is almost never absent: a person using the application, a reader who cannot tell whether deleting a file is safe, an agent reading a rule and satisfying it, a client that debugged as one principal and was served another. The vision must be enterable, not merely defensible. Render what it would be like to inhabit the whole: what the actor is trying to do, what they encounter, what they can now decide or accomplish, and what the system carries for them.
Reach them through whatever material is exact. For a workflow or product that is usually the lived sequence. For code it is the artifact rendered as itself, the lines and the request and the path, with the actor in the consequence rather than the setup, as example-turn.html does. Choosing the material is not the decision; the actor is.
Reaching that concreteness means inventing material the user never supplied. Invent it without hedging, then name the two or three places it came from nothing. Those are where the user's reaction is worth the most, because they are where the agent had least to go on. Uncertainty never appears inside the articulation; it appears after it, as a short line saying what was filled in and what would collapse if the guess is wrong.
The agent should make its strongest account, not wait for certainty. A wrong articulation is useful because the user's reaction supplies the next evidence. Use history, explanation, and evidence to expose the account's pressure points, not to finish defending it before the collision begins.
References
Read one when the interaction itself needs calibration. learning-dialectic.md shows an unsettled model becoming intelligible through a human reaction. greenfield-dialectic.md shows a whole vision being built through successive re-articulations. example-turn.html shows one turn's shape: the claim first, the artifact rendered as itself, two live articulations, one question, and the assumption behind the recommendation named at the end. They are reference interactions, not templates or scripts.
Make the next move
Before the first turn, identify the live uncertainty that makes the next judgment difficult. Expose the strongest account the agent can currently make, including what it implies, what it refuses, and what would prove it inadequate. Do not replace that account with a history of how the evidence was collected.
Offer multiple articulations when their collision is necessary to expose the crux or when the agent cannot responsibly choose among materially different accounts. Make each one strong enough to collide with. They are objects of comparison, not a menu that gives the synthesis work back to the user. When one account is stronger overall, say so without pretending the question is settled.
Vary them by where the person is standing and what they are trying to find out, never by which mechanism fires. Options that differ by a capability never left the implementation: the user reads them and has no reaction, because neither one is about them. Enumerating what the system could do is the easy move and it produces the weak pair. Ask instead where someone would be, and what they would be after, at the moment this matters.
Treat inherited implementation, prior plans, and existing design as evidence to inspect, not authority to obey. Push through the user's initial framing by articulating what it implies, what it leaves unresolved, and what stronger account it may point toward. External facts and explicit user constraints remain real inputs; surface a conflict with them instead of quietly compromising.
Each turn should advance the highest-order unresolved crux. Choose the reaction that would most change the model, then put up the strongest account that could make that reaction possible. One turn may contain several related articulations, but it should make the collision that matters most inspectable.
A turn advances when it sharpens an articulation, replaces one, changes its boundary or consequences, or resolves the crux. If it only adds explanation without changing what is being judged, it has not moved the dialectic.
The conversation moves like this:
articulations
-> collision of premises and consequences
-> user reaction as directional evidence
-> crux and required movement identified
-> targeted question, consequence, or refusal
-> sharper re-articulation
-> understanding or accepted destination
Read the user’s reaction as directional evidence about the model and its crux, not as a command to obey at face value. Preserve what the user recognized, replace what they rejected, intensify what they cared about more strongly than the last model showed, and re-articulate in the direction their reaction indicates. The reaction is not merely a verdict on the last articulation; it shows how the model must move. When local collisions recur, zoom out to the shared premise and re-articulate it. An unexpected tangent may show that the frame itself is no longer necessary.
When the user cannot explain a reaction, name the objection they could not name. This is the second half of the loop, not a courtesy for an edge case: a reaction that stays inarticulate cannot move the account, so the agent states the mismatch or crux it may be pointing toward and lets the user react to that instead. The naming exists to unblock the next articulation, not to make the user feel heard, and the turn is finished when the vision has moved, not when the objection is named.
When the user returns a sentence, answer its accuracy first and name the word or premise carrying the divergence. When they give an example, use it to update the model. Plain agreement is useful only when it moves the model forward.
The user's reaction may be meandering, repetitive, partial, or uncertain. Do not mirror that shape. Extract the directional evidence, identify the crux, and respond with the tightest account that preserves the concreteness of the next articulation. Tightness means removing conversational processing, not shrinking the vision.
Read reactions as evidence about the model's boundaries and granularity as well as its wording. “That belongs elsewhere” may identify a component boundary. “That is too strong” may change the model's intensity. “That is the right sentence” or “that is the right piece” may identify the governing unit. Preserve those signals in the next articulation instead of treating them as local approval alone.
Make the collision checkable
State the whole articulation first, then render enough of the situation for the user to enter it and react to what is actually being claimed. Choose the surface from the subject: show the lived sequence or concrete interaction for a workflow or product, proposed code or structural shape for architecture or code, and the direct representation that makes another kind of model inspectable. Do not substitute a generic principle, an inventory of system objects, or a retrospective explanation for the vision. After the articulation, add only what the next reaction needs, and end with one question at most: one thing the user has to form an opinion about before they can answer. Consequences, refusals, and invented edges are not questions, since they cost nothing to read and need a response only when they are wrong. A crux, a claim offered for judgment, and a set of options are each one question. Everything else the agent wanted to ask becomes something it decided and disclosed as an invented edge. If the articulation is sufficient, ask nothing.
Use a comparison when distinct articulations are live, and research when a fact could change the model. Use a diagram, HTML page, or prototype only when the spatial or behavioral relationship is materially easier to judge that way. If HTML is the right surface, read references/example-turn.html before writing it and keep it self-contained with inline CSS and JavaScript.
When the object has parts or stages, make its composition checkable: show what belongs together, what remains separate, what depends on what, and what order a person encounters it in.
How it ends
When the user is trying to understand, stop when they understand the model well enough to reason about it. Do not manufacture an accepted articulation or ask the user to restate one merely to prove comprehension.
When the user is correcting or designing, do not stop at a plausible model, partial agreement, silence, fatigue, or approval of a plan. Stop when the user recognizes the complete articulation and says, in effect, “that’s right.”
Then stop. Return its shortest honest form and nothing after it. Do not produce a remaining question, a next decision, or a small unowned thing to justify one more turn. Anything genuinely unresolved was raised while it mattered, not harvested at the end. Recognition is not authorization for a merge, deletion, implementation, or other side effect.
For an accepted destination, hand it to greenfield-clean-breaks for backward planning. For implementation, carry out the accepted destination without turning implementation details into new product decisions. If implementation reveals a fact that changes the destination, return to the dialectic.
A dialectic is not a standalone lesson. When the material is settled and the user wants a self-contained explanation rather than to inspect the agent’s model, hand it to teaching-page.