Diagram First

Overview

Answer questions with a short ASCII diagram first, then a concise explanation. Prefer visual structure over prose when describing flows, architecture, or logic.

Default Output Pattern

1. Diagram first (in a code block for alignment).
2. Legend (1–3 bullets, only if needed).
3. Explanation (map diagram nodes/edges to the user’s question; cite key files/functions if relevant).
4. Next check (optional: one quick validation step or a clarifying question).

Choose a Diagram Type

Use the smallest diagram that answers the question:

• Flowchart: branching logic, error paths, validation, “what happens when…”.
• Sequence diagram: request/response, streaming, callbacks, cancellation, distributed coordination.
• Component diagram: modules/services and their responsibilities.
• Pipeline diagram: stage-by-stage compute (tokenization → forward → sampling → decode), incl. distributed ranks.
• State machine: lifecycle of a request, streaming states, tool-call modes.
• Side-by-side: comparisons (“left vs right”, “old vs new”, “A vs B”) using two columns.

Diagram Templates

Sequence (streaming / HTTP)

Client -> HTTP Handler (rank0) -> (broadcast request) -> All Ranks: generate loop
Client <- HTTP Handler (rank0) <- (stream tokens/chunks) <- rank0 only

Flowchart (validation / routing)

request
  |
  v
parse -> validate -> route -> run -> shape response -> send
                 \-> error -> JSON error

Component map (codebase)

api/ (HTTP schemas + shaping) -> model_provider (load) -> tool_fixes (normalize) -> mlx_lm (generate)

Side-by-side compare

LEFT (A)                        RIGHT (B)
---------                       ---------
...                             ...

Practical Rules

• Keep diagrams tight: ~10–30 lines; add a second diagram only if it reduces confusion.
• Include the happy path; add the most important failure/cancel path if relevant.
• Use stable labels (same node name = same thing) and avoid vague boxes (“stuff happens”).
• If you need to assume something, label it clearly in-diagram (e.g., “(assumption)”) and ask a clarifying question.
• Don’t leak secrets/PII in diagrams unless the user already provided it and asked to include it.