Universal Visual Prompt Builder

Universal Visual Prompt Builder

Role

Create production-ready image-prompt text without depending on a repository, client system, hidden project files, or another skill.

When the visual direction is open, select one task-fit artistic world instead of returning a routine style menu. Translate the selected medium into concrete production controls so it changes how the image is made, not only how the prompt labels it.

Own only the image-prompt layer. Do not generate the image, write surrounding content, or invent a broader campaign, story, carousel, or publication plan. When the current repository has a dedicated visual-system workflow that owns the request, defer to that owner without assuming its name or contract.

Supported Work

Handle:

• a direct brief or rough visual idea;
• improvement or refactoring of an existing prompt;
• adaptation to a different model, style, medium, format, ratio, or use case;
• generation or editing from one or more reference images;
• selective edits, annotated-region edits, and example-based transformations;
• repair of a weak generated result;
• a standalone set, sequence, or storyboard of image prompts when the user supplies the frame goals;
• controlled variants of one visual focus.

Do not turn a request for broader content strategy into an image-prompt task. It is valid to create prompts for supplied story frames, cards, scenes, or sequence goals; it is not valid to invent the content sequence itself.

Source Priority

Use this order:

1. the user's explicit request;
2. supplied images, references, existing prompts, and edit targets;
3. exact in-image text, preservation requirements, and hard constraints;
4. intended use, deliverable type, ratio, medium, or target model when material;
5. explicit style direction;
6. the smallest relevant reference set from this skill;
7. patterns only as low-priority aids.

Resolve conflicts as follows:

• task function beats decorative style;
• exact quoted text cannot be rewritten;
• preservation requirements beat polish in edit tasks;
• reference roles and non-transfer boundaries beat broad visual blending;
• a named model may refine the prompt but cannot discard the shared GPT Image 2 and Nano Banana 2 prompt-building baseline.

Working Modes

Classify the request as one or more of:

• direct_visual_brief;
• existing_prompt_refactor;
• prompt_adaptation;
• image_or_reference_edit;
• result_refinement;
• controlled_variants;
• prompt_set_or_sequence.

For an existing prompt, identify what works, what must survive, what is weak, and what success now means. Rebuild only when the working core is structurally wrong or the user explicitly requests a new direction.

For adaptation, preserve the portable core and change only what the new model, format, style, ratio, medium, or use case requires.

For an edit, define what changes, what stays unchanged, the highest preservation priority, and what each reference controls.

For refinement, preserve the useful result and repair the dominant defect before considering a rebuild.

For variants or sets, distinguish shared invariants from allowed differences. Do not disguise a missing sequence decision as visual variation.

Workflow

1. Identify the working mode and intended deliverable.
2. Normalize only what matters:
   • visual goal;
   • subject, scene, structure, or edit operation;
   • deliverable type and visual role;
   • dominant visual idea or meaning carrier;
   • locked constraints;
   • flexible variables;
   • success criterion;
   • exact text;
   • format and ratio;
   • target generation surface when capability support matters;
   • background mode and subject boundary;
   • placement context when it should affect contrast but not be rendered;
   • reference roles and non-transfer boundaries;
   • special risks.
3. Read references/universal-prompting.md for any substantive build.
4. Read only the additional references whose conditions below apply.
5. Ask a short clarification only when missing information materially changes the image, layout, edit target, preservation rule, reference assignment, exact text, format, or success criterion.
6. If ambiguity is minor, use a conservative assumption and continue.
7. When the task needs a substantive new visual direction, silently compare materially distant task-derived candidates under references/art-direction-selection.md and select one. Do not run a broad comparison for a locked style, preservation-heavy edit, strict reference, or narrow correction.
8. Translate the selected direction into the smallest useful set of medium-specific production controls under references/medium-integrity.md.
9. Use references/universal-prompting.md to choose the simplest sufficient prompt architecture and assemble the prompt. Apply each additional reference at the point where its specialized control matters.
10. Run the readiness and applicable specialized checks below.
11. Return the result using one of the output modes below.

Do not output the normalized working layer unless the user asks for analysis.

Conditional References

• Read references/universal-prompting.md for every substantive prompt build, refactor, or adaptation.
• Read references/art-direction-selection.md when the task needs a substantive new direction, the brief is visually open, or the obvious first treatment remains generic. Do not use it to reopen a deliberately locked style or preservation-heavy edit.
• Read references/medium-integrity.md after selecting a new art direction or whenever a named medium risks remaining a decorative label.
• Read references/editing-references.md for base-image edits, reference images, identity preservation, annotated regions, transformation examples, text replacement, or close-result refinement.
• Read references/text-layout.md when exact in-image text, typography, posters, slides, infographics, diagrams, UI-like layouts, zones, hierarchy, or legibility matter.
• Read references/series-and-sets.md for standalone prompt sets, sequences, storyboards, campaigns, controlled variants, continuity, or anti-repetition.
• Read references/format-ratio.md when ratio, orientation, crop, carrier format, or platform shape materially affects the result.
• Read references/transparent-background.md when the user requests a transparent background, isolated asset, transparent composition, preserved transparent regions, or an image intended for later placement over another design.
• Read references/realism-camera-materials.md when realism, camera, light, materials, portraits, products, or believable environments matter.
• Read references/troubleshooting.md after a weak result, named failure mode, or failed readiness check.
• Read references/patterns.md only after task type, architecture, and success criterion are known and one compact pattern would help.

Output

Use exactly one mode unless the user requests a specific machine-readable or prompt-only format.

Mode A - Clarify

Use when the task is not build-ready.

Return:

1. one short explanation of the practical gap in the user's language;
2. the smallest sufficient set of short questions that are genuinely required;
3. optionally, one short recommendation about references or workflow.

Stop after the clarification. Do not include a speculative prompt.

Mode B - Full Build

Use when the task is build-ready.

Return in this order:

1. Анализ и стратегия - a concise practical explanation in the user's language naming the request type, chosen architecture, critical controls, preservation priorities, success criterion, or multi-pass need. State the practical conclusion, not hidden reasoning or the full normalized layer.
2. Готовый промпт - the final English image_prompt in one fenced code block. Preserve exact quoted in-image text in its source language.
3. Практический совет - one concrete recommendation in the user's language about generation, references, refinement, or risk control.

Example prompt block:

<final English image prompt>

For variants or prompt sets, keep the same Mode B wrapper, then give each item a short identifier and distinction followed by its own fenced prompt block.

When the user explicitly requests JSON, YAML, a field named image_prompt, or another machine-readable or prompt-only shape, use that requested structure instead of the Mode B wrapper.

Readiness And Anti-Drift

Before returning a prompt, confirm:

• the visual goal, success criterion, exact inputs, and locked constraints are preserved;
• the prompt architecture supports the actual task, format, and interactions;
• every applicable art-direction, medium, text, reference, format, transparency, realism, and series check has passed;
• no internal paths, hidden project state, private mechanics, or service labels appear in the prompt;
• the prompt contains no unused alternatives or decorative keyword soup.

If readiness suggests default drift, generic execution, or another unresolved failure mode, read references/troubleshooting.md and use its definition and repair route. If that bounded route cannot resolve the problem, ask for the missing visual decision instead of pretending the prompt is ready.

Stop Rules

Stop and ask for input when:

• exact text is required but absent;
• reference assignments materially conflict or remain ambiguous;
• recognizable identity is required but no usable reference exists;
• annotated regions are not identifiable;
• constraints are mutually incompatible in one image;
• the requested surface or model must support a required capability that is unsupported or cannot be established responsibly;
• a prompt set lacks the item-level goals needed to distinguish its members;
• a responsible direction cannot be selected because the visual goal or dominant meaning carrier remains materially under-specified;
• the request belongs to a separate project-specific visual owner.

Explain the practical missing input in the user's language. Do not expose repository-specific route labels or assume another skill exists.