Dialectic

Dialectic owns convergence before a plan exists. The agent presents a positive synthesis; the user reacts freely; the agent treats that reaction as directional data and presents a sharper synthesis. Success is explicit recognition of the destination, not exhaustion, compromise, or a long list of answered questions.

positive synthesis
-> user reaction
-> directional data
-> revised synthesis under pressure
-> explicit acceptance

Compose With

• Use one-sentence-test to compress a fuzzy destination once the model is substantially right.
• Hand an accepted destination to greenfield-clean-breaks when the current system must be worked backward into owner changes, deletion waves, and asymmetric refusals.
• Use grill-me instead when a plan already exists and the job is to interrogate its decision tree.
• Use ui-design for UI-specific product design that must become a buildable interface.

Establish The Frame

Start from the desired product or system, not the inherited implementation. Separate constraints before synthesizing:

Hard constraints:
  user outcomes, safety, security, durable data, external reality, explicit
  promises the user chooses to preserve

Suspended constraints:
  current APIs, names, compatibility paths, package boundaries, prior plans,
  helper shapes, and implementation effort

Inspect the repository when facts affect feasibility or reveal a real product promise. Do not let the current code supply the first model or quietly convert an inherited choice into a hard constraint.

Present A Positive Synthesis

Do not open with a questionnaire or make the user design the answer. Lead with the clearest model you can now see:

Current read:
  What I think the user is reaching for.

Uncompromising vision:
  The cleanest model with inherited constraints suspended.

Why this model:
  The important distinction it clarifies and the value it preserves.

Tradeoffs and refusals:
  What this direction deliberately gives up or makes impossible.

Pressure point:
  The strongest unresolved consequence, counterexample, or fork.

Recommendation:
  The direction I recommend now, stated plainly.

Offer multiple options only when a real fork remains. Recommend one. A menu of equally weighted ideas gives the synthesis work back to the user.

Use the labels only when they improve clarity. After the first round, do not repeat the full template. Keep the exchange conversational: show the revised model, what materially changed, and the next pressure point.

Read Reactions As Directional Data

The user may approve one clause, reject a distinction, challenge a tradeoff, offer an analogy, contradict themselves, or ramble. Extract the signal without requiring them to restate it formally:

Preserve:
  language, outcomes, or principles that produced recognition

Reject:
  assumptions, distinctions, or consequences that felt wrong

Intensify:
  values the user cares about more strongly than the prior synthesis showed

Resolve:
  tensions or ambiguities the reaction exposed

Do not merely paraphrase the reaction. Revise the model, say what materially changed, and show the consequence of that revision.

Revise And Apply Pressure

Each round should do three things:

1. Preserve the parts that earned clear recognition.
2. Replace the weakest assumption with a sharper positive synthesis.
3. Pressure-test one important implication, edge case, refusal, or competing principle.

Keep the live model small enough to hold in the user's head. Prefer one strong distinction per round. Ask a question only when its answer would materially change the synthesis, and still provide your recommended answer for the user to react to.

If the loop stalls, name the exact unresolved fork, state what evidence would resolve it, and recommend a side. A prototype or repository investigation may provide evidence, but neither substitutes for acceptance.

Acceptance Gate

Do not infer convergence from silence, fatigue, partial approval, or the absence of another objection. The destination is accepted only when the user explicitly recognizes the whole model as what they want. Approval with a caveat starts another revision round.

Once accepted, freeze a compact destination artifact:

Accepted destination:
  One concrete product or system sentence.

Mental model:
  The central objects, verbs, boundaries, and owner.

Hard constraints:
  What must remain true.

Refusals and non-goals:
  What the destination deliberately does not preserve.

Consequences:
  The most important tradeoffs the user accepted.

If this cannot stay succinct, run the one-sentence test and continue the loop. Do not begin backward planning while the destination still contains competing models hidden behind "or", "also", "sometimes", or compatibility language.

Transition

After explicit acceptance:

Thinking-only request:
  Return the accepted destination and stop.

Existing-system replacement:
  Load greenfield-clean-breaks, bring the current system back into view, and
  work backward through owner changes, deletion waves, verification, and old
  path removal.

Durable architectural decision:
  Preserve the settled decision and rationale in an ADR at the appropriate
  point in the repository workflow.

Do not implement early to create artificial momentum. Implementation follows the accepted destination unless the user explicitly asks to collapse the design loop and proceed with a stated assumption.

Completion Check

Before handing off, confirm:

Did I present a positive synthesis rather than only questions?
Did the user's reactions materially change or strengthen the model?
Did I expose the real tradeoffs and pressure-test the strongest implication?
Did the user explicitly accept the complete destination?
Can the accepted destination fit in one concrete sentence?
Is the next move clearly thinking-only, clean-break execution, or ADR capture?