Skip to content
Trurlic

Design Conversations

trurlic design runs a structured, Socratic design conversation for a component. Instead of the AI suggesting an architecture, it asks probing questions — forcing you to think through tradeoffs. Decisions are recorded as you answer.

Running a design conversation

Section titled “Running a design conversation”
Terminal window
trurlic design auth

The AI asks questions like “What matters more — latency or durability?” and “How should this handle partial failures?” Each answer produces a decision recorded immediately to .trurlic/decisions/.

Every decision writes to disk as it happens — no batching. If the session is interrupted, all decisions made so far are preserved.

Session modes

Section titled “Session modes”

Fresh (default)

Section titled “Fresh (default)”

Starts a new conversation from scratch for the component. The workflow engine assesses which concerns are uncovered and guides the conversation accordingly.

Terminal window
trurlic design auth

Continue

Section titled “Continue”

Resumes a previously interrupted session. Picks up where you left off.

Terminal window
trurlic design auth --continue

Revisit

Section titled “Revisit”

Re-examines existing decisions for drift or staleness. Useful when requirements have changed and you want to challenge prior choices.

Terminal window
trurlic design auth --revisit

Task-focused design

Section titled “Task-focused design”

Focus the conversation on a specific task:

Terminal window
trurlic design auth --task "add OAuth2 PKCE flow"

The workflow engine scopes concern coverage to the task context, skipping areas that are irrelevant.

Provider and model

Section titled “Provider and model”
Terminal window
trurlic design auth -p anthropic -m claude-sonnet-4
trurlic design auth -p openai -m gpt-4o
trurlic design auth -p openrouter
trurlic design auth -p gemini -m gemini-2.5-flash
trurlic design auth -p ollama -m llama3.1
trurlic design auth -p custom -m my-model

See Configuration for provider resolution, API key setup, and the full list of supported providers.

How it works

Section titled “How it works”
  1. trurlic design resolves the LLM provider and API key.
  2. The workflow engine (advance) computes the first step from the current graph state.
  3. The session module calls the LLM with the step prompt and component context.
  4. The LLM asks a Socratic question. You answer.
  5. The session module extracts decisions from your answer and records them immediately.
  6. After recording, the in-memory graph cache is rebuilt to keep it consistent for subsequent writes in the same session.
  7. advance computes the next step. Repeat until all concerns are covered.

The session module is the only module that calls LLM APIs. The workflow engine and store are pure — they never touch the network.

Design principles

Section titled “Design principles”

Questions should make you think. “What matters more — latency or durability?” not “I suggest Redis.” The AI is a facilitator, not a decision-maker. Every architectural choice in the graph came from a human answer to a Socratic question.

After each answer, the graph is immediately consistent and queryable. If you stop mid-conversation, every decision recorded so far is available to coding agents.