SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skills100x-localize

100x-localize

Writing & research

View repository

Updated 10-15 23:45

About this skill

100x-localize

One-line positioning

Input a piece of source copy (any source language), output a default localized version in Mexican Spanish—not a word-for-word translation, but actively compressing length and matching the actual intensity distribution of real corpora. It belongs to the L2 creative generation layer of the 100x system, corresponding to the "2c ad copy localization" step.

When to trigger

The user says:

  • "ad copy localization" / "localize into Spanish" / "translate into Spanish copy" / "rewrite into Mexican Spanish" / "how should this copy be localized"
  • "localize this ad copy" / "translate this script into Spanish" / "adapt this for the Spanish market" / "Mexican Spanish localization"
  • Or directly provides a piece of source copy (any source language, usually English or Chinese copy), requesting a Spanish localized version

Input

Minimum input (Category A, mandatory):source_script—Complete source copy as plain text (any source language). This skill does not accept video/audio files, only text.

Soft supplements (Category B, all have explicit default values, see workflow.md Phase 1）：target_region (default "mx", only accepts mx/generic-latam, other values are downgraded to generic-latam and recorded in meta.warnings ),register_profile(default "default", matching real corpus intensity, not a conservative setting, see "Core Constraints" axiom 2).

Optional upstream outputs (Category C, silently skipped if missing, non-blocking, and never require the user to run other skills first): 100x-persona character/scene cards,100x-search-query search term results.

Output

For structure see schema.json：source_script(source text echo) + target_region + register_ profile + localized_script(Spanish localization output) + meta(required:generated_by、 warnings; optional information fields:compression_ratio）。

Core constraints (4 axioms, see details in axioms.md）

  1. Compression axiom: the Spanish output character count must not exceed 1.10 times the source copy character count (nor be lower than 0.5 times)—calibrated based on the transcription length distribution of the reference corpus; Spanish is naturally longer when expressing the same amount of information, and literal translation will systematically exceed the voiceover duration budget, so active compression and rewriting are required, not word-for-word translation (the two specific numbers 1.10/0.5 are engineering judgments, see details in axioms.md axiom 1)
  2. Register intensity default axiom: by default, match the actual intensity distribution of real corpora; the conservative prohibited-language list in specification document No. 07 (07_西语口播风格规范.md) §5 is downgraded to optional setting profiles/compliance-conservative.md, not the default (that document's §1 itself states it is "not a practical conclusion induced from Spanish viral samples in the library," so this skill therefore does not treat its conservative recommendations as default verification conclusions that must be followed)
  3. Person consistency axiom: use tú address form consistently throughout, and do not mix in usted/vosotros(known limitation: currently only supports mx/generic-latam the two target_region, and does not support Spain formal usted/vosotros register or Argentine vos conjugation—this is a scope limitation; requesting such regions will be downgraded to generic-latam with a notice, and will not pretend to support them. In addition, the reference corpus actually contains cases where usted appears and is mixed with tú —this axiom is the quality line this skill sets for its own output, not a description of the current state of real corpora, see axioms.md Axiom 3. The current mechanism already covers the four types of bypass techniques: all-uppercase/deaccenting/zero-width invisible character insertion/combining accent insertion (the handwritten validation layer performs double normalization of invisible characters + combining accents for Axioms 2/3, and the authoritative word-family detection of Axiom 4 likewise performs invisible character stripping). Known remaining gaps: schema.json layer's ajv pattern still has no preprocessing capability (a structural limitation of the JSON Schema specification), and the handwritten layer's invisible character code point list (6) is not an exhaustive enumeration of Unicode format character categories, see axioms.md TODO for Axioms 2/3/4)
  4. Anti-fabricated-authority-claim axiom: must not introduce authority/certification-type claims not present in the source copy (closed-set word families: FDA/ Harvard/OMS/clinical-validation-type expressions) — known limitation: this can only block keyword-level additions within the closed-set word families (normalization has been done for written variants such as deaccenting, mixed case, and dotted-letter abbreviations, but it still cannot block fabricating a fictional institution name/person name not in the closed set), see axioms.md Axiom 4

Three-phase process

See workflow.md: Phase 1 receive + validate (categories A/B/C,target_region/ register_profile degraded fill) → Phase 2 (register strength decision → language naturalization rewrite → length compression → sentence-by-sentence self-check → batch self-check) → Phase 3 user-triggered rework (L1 single-round compression / L2 tier or region switch / L3 full redo;person register has no corresponding rework path—v1 only supports tú, see "Core Constraints" Axiom 3).

Independent invocation guarantee

Category A (source_script) missing or the user sends a video/audio file → fixed refusal script to ask follow-up questions, no guessing on their behalf, no writing scripts on their behalf. Category B (target_region/register_profile) missing → directly use the documented default values to continue running, no follow-up questions, no blocking. Category C (upstream 100x-persona/100x-search-query output) missing → silently skip,never require the user to run other skills first。

Banned words

let me / hope / perhaps / probably / maybe / possibly / let's; AI customer-service tone (as an AI / I'd be happy to / feel free to);register_profile=="compliance-conservative" the specific banned-word list under this tier is in profiles/compliance-conservative.md; do not fabricate effects/certifications/user testimonials not present in the source copy (see Axiom 4, known limitations above); do not hardcode any customer category dictionary; do not pretend to support regions/ person registers not implemented in v1 (see known limitations of Axiom 3).

Routing (recommended, not a hard dependency)

  • Already have character/scene settings and want the localized tone to fit a specific persona → character-scene-type skill (if already built, its output can serve as Category C reference)
  • Have not yet decided on search keywords and want to find reference material first before localizing → search-keyword-type skill (if already built)
  • After localization, want to generate voiceover/subtitle timelines → voiceover/subtitle-type skill (if already built)
  • Want to return 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; when the corresponding skill has not yet been built, it does not affect this skill's independent work.

Factory self-check

Run node scripts/validate.js --selftest to validate.schema.json The structural-layer constraints declared byrequired/additionalProperties/enum/pattern, including the two conditionable if/then+pattern checks for axiom 2/3) are handled by scripts/validate.js within ajv(a real JSON Schema draft-07 validator,package.json declared devDependency, first npm install run again) actually executed. Axiom 1 (compression ratio) and Axiom 4 (anti-fabricated authority) are two types of cross-field comparison rules that vanilla JSON Schema cannot express structurally (it cannot compare the length or content of two string fields), so they are supplemented by the handwritten code in 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

Source

This skill is built on the following materials:07_西语口播风格规范.md(the direct source of register/naturalization rules, whose self-declaration of "not being an empirical conclusion" is preserved as-is) + reference corpus (providing data support for the compression axiom and the register default axiom)

  • T7 module library CSV (providing structural reference for the module skeleton). For per-file rewrite points and criteria explanations, see sources.md。