smthemex/comfyui_sensenova_u1/prompt-enhancement-skill

SenseNova U1.5 Prompt Enhancement

Turn the brief into one concise prompt for the image generator. Preserve the user's semantic contract, then actively resolve open visual variables so the result feels designed rather than merely complete.

Read references/compact-scene-compiler.md completely before enhancing a brief. Also read references/retrieval-and-reference-compiler.md completely when the brief contains current or source-dependent facts, the user asks for research or references, supplied images must guide the result, or searched visual references would materially improve accuracy or art direction. Read references/core-render-json.md before returning the prompt. Its compact core plus category-specific structure extension is the required output contract. Examples in other references are illustrative and do not change this contract.

Core contract

Silently separate the brief into five kinds of information:

• Locked creative and structural requirements: supplied subject, action,

entities, non-factual visible copy, count, relationship, fixed layout, palette, format, exclusions, and explicitly fictional or counterfactual premises. Preserve these exactly.

• Provisional real-world assertions: supplied locations, routes, dates,

values, rankings, scientific relations, institutional relationships, and label-to-object mappings whose correctness determines a reality-based artifact. Verify consequential claims and relationships before rendering; preserve departures from reality when fiction, speculation, satire, or a counterfactual treatment is explicit.

• Visible-copy contract: preserve explicit render-intended strings exactly,

route factual literals through verification, and resolve genuinely required functional copy for content-bearing artifacts into short final literals. Store every render-intended string in content; never allow typography, layout, structure, or decoration to introduce a string that is absent from content.

• Open visual variables: viewpoint, crop, pose or object state, generic

wardrobe and props, foreground and background depth, lighting, color roles, material response, typography character, spacing, overlap, rhythm, and production finish. Resolve enough of these decisively to author the image.

• Unsupported claims: unsupplied brands, real identities, current or

source-specific statistics, prices, dates, certifications, endorsements, contacts, private data, or claims attributed to an unavailable source. Retrieve reliable evidence when the claim is required and retrieval is permitted and available; otherwise do not invent it.

Fidelity means preserving locked intent, not leaving open visual variables underspecified.

Prompt-to-image text isolation

The enhanced prompt is consumed by an image renderer, not displayed as a design specification. Prevent the renderer from turning prompt prose, JSON field names, or generation rules into visible copy:

• content contains final render-intended literals only. A literal is text

that should be readable in the finished image exactly as stored. Never put a scene description, layout instruction, component description, rationale, prohibition, URL to inspect, source note, placeholder, field name, or generation command in content.

• Describe visual objects outside content without quoting the description as

copy. For example, encode a red banner as shape, position, and material; do not create a content item saying "top red rounded banner".

• Never copy the user's request sentence into theme,

main_subject.description, a title, or a structure member. Compile it into observable image state.

• Every descriptive value must be brief-specific and pixel-causal. Do not emit

reusable policy prose such as "keep all locked subjects", "dominant, supporting, tertiary", "content literal", "no extra glyphs", "the model may solve", or explanations of what a schema field means.

• Do not make a rule list, audit panel, specification rail, legend, footer, or

text block out of composition rules. Leave composition_rules empty by default and encode local constraints in their dedicated fields first. For a single-scene authored result, include at most one short image-wide invariant. When structure is present, include up to three short, independent image-wide invariants only if exact count, continuity, or mapping spans multiple fields and cannot be expressed unambiguously within structure or another dedicated field. Do not mention the user, prompt, model, requirements, fields, literals, or rendering process.

• Keep negative_prompt short. It names visual failure classes only and must

not restate the schema, content inventory, or exact-copy protocol.

• Structure member identifiers are machine-facing and semantically neutral

(step_1, panel_1). Their states describe visible imagery, not captions. Put a step name in content only if the finished artifact genuinely needs that visible label.

• For a text-free result, keep content empty, set typography to

none; text-free image, do not place quoted candidate copy in descriptive fields, and avoid long sentence lists that resemble body copy.

• For a text-free result, make type a short category token rather than a

title-like phrase (illustration, photograph, storyboard, product render, diagram). Keep theme to a short non-imperative subject state, never the request sentence or a deliverable title. Put the resolved visual specifics in main_subject and the scene fields.

Before returning, run a silent candidate-copy audit across every string in the JSON. Ask: "If this exact string appeared legibly in the image, would the user recognize it as requested final copy?" If no, ensure it is not in content, shorten it to concrete visual state, and remove metalinguistic phrasing. Then scan for schema terminology, field labels, hierarchy boilerplate, preservation instructions, model-facing permissions, prohibitions, and rendering-process language. Rewrite such phrases as observable visual state unless the user explicitly requested them as visible copy.

Text-safe compact mode

Use this mode by default for SenseNova U1.5 and other image generators that may interpret descriptive prose as visible text. Long structured prompts can be mistaken for the body copy of a poster or specification board even when content is correct.

• Target 900–1800 serialized characters excluding user-supplied visible copy.
• Keep each non-content string to one short visual phrase. Do not write

explanatory sentences, semicolon chains, policy lists, or rationale.

• Leave composition_rules empty by default. Use at most one rule for a

single-scene authored result, or up to three for a structured result only when indispensable count, continuity, or mapping constraints span multiple fields. Keep each rule under 30 CJK characters or 60 Latin characters.

• Keep negative_prompt under 40 CJK characters or 80 Latin characters and

name only two to four concrete failure classes.

• Keep each layout_composition value under 40 CJK characters. Describe

geometry, not reading-process prose.

• Keep typography to the minimum appearance needed for whitelisted copy. Never

mention schema paths, exact-copy protocols, or regions that should contain no text.

• Use no proposal-board specification rail, material schedule, dimension

schedule, safe-area note, grid measurement, margin value, or production annotation unless the user explicitly requests that item as visible content.

• In structure, keep each state to one compact observable scene. Never add

process commentary or labels merely to explain the member.

Workflow