title: "Default Translation Flow: Concise vs. Detailed",

summary:

"How DevSpeak turns input, tone, and output format into concise, balanced, or detailed generation behavior.",

category: "core-translation-engine",

tags: [

"default-flow",

"verbosity",

"prompt-branching",

"refinement",

"generation-modes",

],

lastReviewed: "2026-04-08",

};

Default Translation Flow: Concise vs. Detailed

DevSpeak presents a simple translation workflow in the UI, but the underlying runtime makes a few important decisions based on tone, attachments, and whether the request is a first-pass translation or a refinement.

What This Page Covers

  • How the app interprets concise, balanced, and detailed requests.

  • Where the prompt builder branches internally.
  • Which stages are shared by translation, refinement, and file-backed workflows.

    • Flow Overview

    1. The user enters source text and chooses audience, context, output format, and tone.

    2. The server builds a prompt using the current settings and optional custom instructions.

    3. The provider layer resolves the managed path or a valid BYOK configuration.

    4. The response is sanitized, rendered as markdown, and persisted with metadata.

    5. If the user refines the result, DevSpeak reuses the previous output plus explicit feedback instead of starting from scratch.

    Concise, Balanced, and Detailed

    The tone control is a continuous slider, but its behavior changes at the midpoint rather than gradually. Anything below 50 produces a tightened rewrite; 50 and above produces a full document.

    | Mode | Tone Range | Runtime Behavior | Best For |

    | :------- | :--------- | :------------------------------------------------------------------------------------------------ | :------------------------------------------------ |

    | Concise | 0-49 | Rewrites and tightens the original request without inventing major new sections | Fast drafts, cleanup, quick technical reframing |

    | Balanced | 50 | Uses the detailed prompt path, but typically with moderate user input size and standard structure | Default day-to-day feature work |

    | Detailed | 51-100 | Expands the request into a fuller engineering document with assumptions and structured sections | RFCs, implementation-ready specs, deeper analysis |

    Because the change is a step and not a ramp, moving the slider from 49 to 50 is the single most noticeable adjustment you can make. If a result is longer or shorter than you expected, check which side of the midpoint you are on before changing anything else.

    Refinement works differently: it takes your original input, the document already generated, and your feedback together, then returns a complete replacement document rather than a patch.

    Shared Stages Across Flows

  • First-pass translation: Uses the current translation settings and returns a complete response.

  • Refinement: Reuses the existing output and applies explicit change requests from the user.
  • File-assisted translation: Keeps the same output controls, but attachment handling and document extraction happen before generation.

    • Operational Notes

  • Fast AI mode can trigger a lighter-weight generation loop for draft feedback while the user types.
  • File-backed translations and document jobs generally prioritize compatibility and completeness over the fastest path.
  • Every completed translation can be persisted and reloaded later with its settings, attachments, and history metadata.