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

100x-segment

Agents & models

View repository

Updated 10-15 23:45

About this skill

100x-segment

One-line positioning

Input a plain-text TikTok/UGC voiceover script (English or Spanish), output three independently layered segmentation results: L1 paragraph logic (10 modules + 7 archetypes), L2 shot-purpose prediction, L3 inline breath-mark markers. Belongs to the 100x system "1 segmentation" step.

When to trigger

User says:

  • "help me segment" / "how should this script be segmented" / "break down this voiceover copy" / "where should the breathing pauses go in this passage"
  • "segment this script" / "break this voiceover into beats"
  • Or directly give a plain-text voiceover script and ask to "split by module" / "mark the breath pauses" / "how should this be cut"

Input

Minimum input (Category A, strictly required):source_text——a plain-text voiceover script/copy.This skill does not accept video files/video frames, it only takes text (see workflow.md Phase 1）。

Category B (soft input): none——source_text itself already contains all the signals needed for segmentation; there is no need to supplement selling points/audience/tone before it can start working (consistent with the 100x-persona structure, different from 100x-search-query , see workflow.md"Category B explanation").

Language scope (v1 explicit boundary, not a degradation):language is automatically determined from source_text ; only English or Spanish is accepted; when it is determined to be another language, or English and Spanish are mixed to the point where the primary language cannot be determined, reject directly—do not guess, do not do it on the user's behalf.

Optional upstream output (Category C; when missing, skip silently, do not block, and never require the user to run another skill first): This repository currently has no skill output that can serve as upstream product profile/topic material.

Output

For the structure see schema.json：source_text/language/archetype(7 archetypes A-G, single or "primary + secondary" composite such as "G+E"）/segments[](each item contains segment_id/module[10-enum]/ raw_text/text_annotated[inline marker version with ‖strong breath pause·weak breath pause]) /shots[] (each item contains shot_id/segment_refs[]/shot_purpose[7-enum standard definition]) /meta. Optionally render one more human-readable Markdown paragraph table.

Core constraints (4 axioms, see axioms.md）

  1. segments[].module locks the 10 enum (11 modules minus Localization, which runs through the whole text) + archetype locks the 7-archetype format (single or "primary + secondary" two different letters)
  2. Breath pauses are inline markers, only added, never subtracted——text_annotated after removing ‖/· must be verbatim identical to raw_text ; no deleting/adding/changing words is allowed
  3. The criteria for strong/weak breath pauses are locked down (conjunction-at-start-of-turn > end-of-sentence punctuation/em dash > comma/semicolon; one breath-force split at EN 14 words/ ES 16 syllables), judged by fixed priority, no free improvisation (known limitation: EN so simultaneously has two uses, "contrastive/causal conjunction" and "degree adverb/emphatic word" (so much/so many); pure keyword matching cannot distinguish them, and only the most common collocation "so much"/"so many" has been excluded; other emphatic uses may still be misjudged as contrastive conjunctions and forcibly require a strong breath pause——this is ahard fail criterion rather than a soft warning; the cost of misjudgment is higher than the following one; the two rules "not marking a breath pause after a filler word" and "not marking one before a price number" are heuristic keyword matching, not true semantic judgment,like the preposition/filler ambiguity will cause a certain false-positive rate, only a warning is issued and it does not fail, see axioms.md Axiom 3 TODO; the end-of-sentence punctuation criterion uses a fixed, non-exhaustive abbreviation dictionary (a.m./p.m./Mr. etc.) to exclude common abbreviations periods, has been confirmed fixed on real English corpora for these specific abbreviations, but abbreviations outside the dictionary will still be misjudged as sentence ends; in addition, sentence-end punctuation stuck to the next word with zero space (missing space) currently only for !/? is judged as a missing breath point,. out of consideration for avoiding conflicts with abbreviations/decimal points, still requires a real whitespace character, zero-space stuck . does not trigger the criterion for now, all see axioms.md Axiom 3 TODO)
  4. shots[].shot_purpose Lock 7 enumerates standard definitions, and must fully cover segments[](reference integrity + zero orphans, known limitation: this skill is plain-text input, shots[] only pre-judges shot purpose, does not include time_bucket/visual_description/ camera_language/audio_plan——these fields depend on the actual video footage, outside this skill's scope)

Applicability boundary: the 11 modules/7 prototype skeleton is taken from the TikTok US DR supplement oral script blueprint, applicability to other copy structures (narrative-oriented/interview-oriented and other non-DR private-domain sales script types) has not been verified.

Three-phase process

See workflow.md: Phase 1 receive + validate (category A + language range validation, no category B, category C silently skipped) → Phase 2 (2A paragraph segmentation → 2B shot purpose pre-judgment → 2C breath point annotation → 2D self-check S1-S8 + rendering) → Phase 3 user-triggered rework (L1 paragraph layer/L2 shot layer/L3 breath point layer three-level granularity, directly corresponding to the three-layer cut points).

Independent invocation guarantee

Category A (source_text) missing → fixed refusal wording to follow up, no guessing on behalf, no accepting video files as a substitute. Language determined to be outside the English-Spanish range → fixed refusal wording, no guessing, no routing. Category C (upstream output) missing → silently skipped,never require the user to run other skills first。

Forbidden words

let me / hope / perhaps / probably / maybe / might; AI customer-service tone (as an AI / I'd be happy to / feel free to); do not fabricate module content/effects/certifications/testimonials not in the script original text; do not hardcode any customer category dictionary;text_annotated deletion or modification of raw_text any word is not allowed (see Axiom 2).

Routing (recommended, not a hard dependency)

  • Already segmented, want to assign characters/scenes to each paragraph → 100x-persona（pairings[].segment_ref can backfill this skill's output segments[].segment_id）
  • Want to pair the segmented paragraphs with real shot footage/camera movement/audio (requires video input) → a future publicly repo-ized storyboard-type skill (if already built)
  • Want to first find visual reference images/search keywords → 100x-search-query
  • Want to return to the entry point for re-triage → 100x system main entry (if already built)

All of the above are "recommended on demand after output," not prerequisite dependencies; when the corresponding skill has not yet been built, it does not affect this skill's independent operation.

Factory self-check

Run node scripts/validate.js --selftest to validate.schema.json The structural layer constraints declared by (enum/pattern/additionalProperties/minItems/uniqueItems etc.) are scripts/validate.js within ajv(real JSON Schema draft-07 validator, package.json declared devDependency, first npm install then run) actually executed, not handwritten field enumeration. Axiom 2 (text_annotated/raw_text cross-field comparison), Axiom 3 (breath point strength priority + one-breath upper limit, requires parsing the string into segments and reasoning segment by segment), Axiom 4 (shots/ segments cross-array reference integrity + zero orphans + sequential numbering) are criteria that the four types of vanilla JSON Schema cannot express structurally, supplemented by 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 for the scenario of "plain-text spoken-word script segmentation and breath-mark annotation," and includes paragraph logic (10 modules + 7 archetype system), shot-purpose prediction (7 standard purpose categories), and an inline breath-mark rule system (Axioms 2/3). For detailed specifications of each criterion and field design, see sources.md。