100x-visual-fission

100x-visual-fission

One-line positioning

Input ≥2 reference materials of the same series of characters/scenes (reverse-engineered JSON or text-degraded description) + product copy, output a media fission prompt matrix that locks character/scene/product identity: choose one of four media structures (single frame/first-last/first-middle-last/visible results in days) × VTP original A/B/C camera presets, with fixed negative prompt constants attached. Belongs to the L2 creative generation layer of the 100x system, corresponding to the "5 visual fission" step.

When to trigger

User says:

"Help me fission this visual reference" / "How many variants can this product produce" / "Generate a fission prompt matrix" / "media fission" / "visual fission"
"fission this reference" / "generate prompt variants for this product" / "multiply this reference image into variants"
Or directly provide a set of already reverse-engineered character/scene JSON (or text description) + product copy, requesting a set of fission prompts

Input

Minimum input (Category A, hard required, both items required):references[]——At least 2reference materials of the same series of characters/scenes, which can be completed VTP reverse-engineered JSON (vtp_prompts_01/02 schema), or a text description degradation when there is no real reference image (must be truthfully annotated in the produced source_material_note );product_brief—— product name + category + effect narrative text. Fewer than 2 references cannot perform "extract commonalities" (cannot distinguish what is changing), and will be rejected, see workflow.md Phase 1 Category A validation.

Soft supplements (Category B, missing uses default values, see workflow.md）：aspect_ratios_wanted(default ["9:16"]）、variant_count(default 2, this is a parameter freely specified by the user in VTP step 04, not a structural criterion of this skill).

Optional upstream outputs (Category C, missing silently skipped, does not block, and never requires the user to run other skills first): 100x-persona produced PersonaSceneBundle(if it already exists, directly reuse its identity anchors and skip reverse engineering).

Output

Structure see schema.json：constants(character/scene/product three anchors) + variables(commonality/variable extraction of VTP step 03) + media_plan(choose one of four media structures + frame plan) + prompt_sets(N groups of prompts generated in VTP step 04) + fission_variants(camera fission of VTP step 06, at least one per frame) + negative_prompt (fixed constants of VTP step 07) + meta. Optionally render a human-readable Markdown summary (media structure + frame list + fission entries corresponding to each frame).

Core constraints (4 axioms, see `axioms.md`）

The character/scene/product three anchors are quantitative and must appear verbatim in every fission branch (the substring inclusion check is a case-insensitive literal match, not a semantic equivalence judgment—two passages with different wording but describing the same person, or two different products that happen to share the same generic boilerplate, may both cause this check to produce false negatives/false positives, see axioms.md Axiom 1)
The media fission axis (single frame/first-last/first-middle-last/visible results in days) is superimposed on top of the VTP original three-tier axes of camera/lighting/color grading, not a replacement——this is the core difference between this skill and the original VTP design (known limitation: media structure determination requires multi_day the reason for the branch must cite a number, but the machine cannot verify whether this number faithfully reflects the input narrative, see axioms.md Axiom 2 TODO)
Each frame plan must actually be rendered by at least one fission product, and the source referenced by the fission product must actually exist—planned but not rendered, or rendered but pointing to a nonexistent plan, both are judged FAIL
Each final prompt fixes the negative-word constants + a realism fallback suffix; do not copy non-visualizable/exaggerated-claim source sentences into the image Description (known limitation: the non-visualizable claim keyword table has so far only been validated on real English + Spanish corpora in the health supplement category, and expanding across more categories will most likely require extending the table, see axioms.md Axiom 4 TODO)

Three-phase workflow

See workflow.md: Phase 1 intake + validation (category A/B/C + non-visualizable claim pre-check) → Phase 2 (VTP seven-step skeleton: 01/02 reverse-derive or downgrade → 03 extract commonalities → medium-structure determination (new in this skill) → 04 generate N prompt sets → 06 camera-position fission → 07 fix negative words → item-by-item self-check) → Phase 3 user-triggered rework (L1 single item / L2 single-frame or structural level / L3 all).

Independent-invocation guarantee

Category A missing (references[] fewer than 2 items, or product_brief unavailable) → a fixed refusal script asks follow-up questions, without guessing on the user's behalf or forcing one reference into "2 items". Category B missing falls back to default values + a meta.warnings notice. Category C (100x-persona output) missing is silently skipped,and the user is never required to run 100x-persona。

Banned words

let me / hope / perhaps / roughly / possibly / maybe / let's; AI customer-service tone (as an AI / I'd be happy to / feel free to); ineffective aesthetic words attractive/beautiful/stunning/serene/pristine/elegant/ cinematic/professional/studio/flawless (a direct extension of VTP's anti-studio bias); do not copy non-visualizable/exaggerated claims ("clinically proven"/"cellular health"/"guarantee", etc.) verbatim into the image description (Axiom 4); do not fabricate facts not present in references[]/product_brief ; do not invent specific brand/client names; negative_prompt fixed constants may not be rewritten by the user or the rework workflow.

Routing (recommended, not a hard dependency)

No reference images yet, want to find visual references first → search-keyword-type skill (100x-search-query, if already built)
Character/scene not decided yet, only a piece of script copy → persona/scene-type skill (100x-persona, if already built, output can be fed directly into this skill's category C)
Want to turn fission results into a formal storyboard script/timeline → storyboard-type skill (if already built)
Want to go back to the entry point for re-triage → 100x system main entry (if already built)

All of the above are "recommend on demand after output", not prerequisites; if the corresponding skill has not yet been built, this skill still works independently.

Factory self-check

Run node scripts/validate.js --selftest to validate. The cross-item aggregation rules in Axioms 2/3/4 (frame array length, frame order, referential integrity, zero orphans, negative-word consistency across branches, non-visualizable claim scan) schema.jsoncannot be expressed (in vanilla JSON Schema); what actually performs the判定 is scripts/validate.js：

npm install                                                   # 首次使用先装 ajv
node scripts/validate.js <bundle.json> [bundle2.json ...]     # 校验产出
node scripts/validate.js --selftest                           # 跑内置回归用例（16 项）

Source

The methodological skeleton of this skill is a methodological rewrite of the VTP (Visual Template Proliferator) seven-step pipeline: 《Reverse-Engineering Methodology · Four Types of Reverse Engineering and the VTP Seven Steps》 + reverse-engineering schema (person/scene) + extract-commonalities/generate-variants/translate/fission prompt + the 07 negative-word constants in the methodology document. "Medium fission" (the single-frame/head-tail/head-middle-tail/visible-results-in-days four-choice medium- structure axis, layered on top of VTP's original camera-position/lighting/color-grading axes) is this skill's original extension, not a layer that existed in VTP's original design; for the specific differences see axioms.md opening and Axiom 2. For per-file rewrite points, disclosure of original judgments, and corpus test records see sources.md。