界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
效率与协作 · SKILL.md

flux-3-prompt-doctor

flux-3-prompt-doctor

【效率与协作】Use when diagnosing a FLUX 3 brief before generation. Resolve missing, conflicting, or schema-changing requirements.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: flux-3-prompt-doctor description: Use when diagnosing a FLUX 3 brief before generation. Resolve missing, conflicting, or schema-changing requirements. metadata: author: Black Forest Labs version: "1.0.0" tags: flux, flux-3, bfl, prompt-doctor, validation, routing

FLUX 3 Prompt Doctor

Resolve the decisions that change routing, payload, or production method before a brief becomes a prompt. Ask only blocking questions; use documented defaults or label assumptions for everything else. Preserve the user's product names, exact copy, source roles, and hard constraints verbatim. This skill does not write the final prompt or call the API.

Route by what must survive

Every request names its mode and carries the matching media field:

RequirementRoute
Generate the whole clip from wordsmode: "t2v", no media
An image must appear on screen exactly as shotmode: "i2v", keyframes
Bridge an exact opening and closing framemode: "i2v", two keyframes
Pass through visual waypointsmode: "i2v", keyframes as a storyboard
Continue from an existing endingmode: "v2v", start_video
Render an approved draft without replanningmode: "draft_enhance", draft_cache

Two questions separate the image routes: must these exact pixels be on screen (keyframes)? Can the source simply be described (then attach nothing)? There is no field that carries a subject's identity without putting the source on screen; a brief that needs one either opens each shot from a keyframe containing the subject, or is a REVISE.

Feasibility

  • Action budget: one subject action and one camera move per short clip. Revise briefs demanding many sequential actions, several locations, exact mechanisms, or multiple full dialogue lines.

  • Shot structure: state one continuous unbroken take, or an explicit sequence:

    SHOT ONE: wide aerial of a desert highway at dawn, a single red car speeding through.
    HARD CUT. SHOT TWO: interior close-up, the driver's hands drumming the wheel to the radio.
    HARD CUT. SHOT THREE: from the roadside, the car shrinking into the heat haze.
    Warm engine hum under one continuous music bed across all three shots.
    

    Consecutive shots must contrast hard (scale, location, color) or the cut blends into a continuous take; each shot needs its own beat; one music bed can run across cuts.

  • Contradictions to catch: locked-off camera plus tracking; extreme close-up plus establishing view; exact frame preservation plus major geometric change; one continuous shot plus hard cuts; production-critical generated typography with no post fallback; guaranteed speaker identity across separate generations.

  • Reserve for deterministic post when exactness is a success condition: typography, subtitles, logos; frame-accurate sync; final mixing; precise mechanisms; cross-generation speaker identity. Generation creates footage and causal intent, not a compositor or final mix.

Verdict

Return exactly one state, with a reason, next action, and a compact handoff (templates/brief.md when it should be reusable):

  • READY: route, assumptions, handoff. Recommend a draft first when the concept is unproven.
  • NEEDS INFO: only the blocking questions.
  • REVISE: name the conflict and the minimum repair.

Non-blocking warnings go under Production risks, not a fourth state. Field names and limits belong to the API reference; flag intent conflicts and leave schema validation to flux-3-generate.

我的收藏