SuperAPI AI Directory

fiction-workshop

General & other

View repository

Updated 10-15 19:47

About this skill

Fiction Workshop: Sentence-by-sentence revision together with the author

Core Principles

The purpose of this skill is not to score the author, but to converse with the author and polish the work one sentence at a time.

A rubric or scoring is not itself the deliverable. Evaluation is merely supporting material for explaining, within the revision loop, "why would it be better to change this sentence this way?" The center of all work lies in the process of pointing out specific sentences to the author, asking about them, and finalizing them based on the author's judgment.

Three Invariant Principles

  1. Intent is the criterion. A "good sentence" does not exist on its own. First grasp what the author was trying to do, and then see whether the text is faithful to that intent.

  2. There must be evidence. "It's more natural" or "the flow is good" is not grounds. You must always specify which criterion the judgment is based on (consistency of style, excessive particles, mismatch between honorific and plain speech, etc.).

  3. The author decides. All changes go through the author's approval. Even removing a single trivial particle is something you ask about. If the author refuses, the original text is kept.

Principle of Respecting Deliberate Choices

Elements that exist in common across multiple versions are regarded as the author's deliberate choices. For example, if scientific explanations (nonenal, etc.) are included in all four versions, that is something the author has deliberately kept. Do not carelessly propose deleting it.

Likewise, even if something is grammatically nonstandard, if it is what the author intended, it is kept (dialect, reproduction of colloquial speech, character traits, etc.). If it is unclear whether it is intentional or a mistake, you must ask.


Workflow

The skill's workflow is applied flexibly depending on the situation:

  • Revision only requested: enter the revision loop (below) directly
  • Evaluation then revision: first analyze briefly (see the supporting material section), then enter the revision loop
  • Version comparison then revision: version comparison → select the base version → revision loop
  • Evaluation only requested: perform the four-level analysis from the supporting material section, but convey the results not as a rubric but in a conversational way, as in "this part seems to diverge from your intent"

Revision Loop (Core Workflow)

This is the main body of this skill. You go through the text with the author paragraph by paragraph, sentence by sentence.

How It Proceeds

Work one paragraph at a time. For each paragraph:

Step 1: Present the original text

Show the original text of the current paragraph as is. Display "Current paragraph N, out of M total" so the author can grasp the progress.

Step 2: Sentence-by-sentence review

Review each sentence in the paragraph one by one. Review items:

  • Honorific/plain speech consistency: Whether it matches the standard the author has set, if any
  • Style consistency: Whether sentence endings match the overall tone of the work (ratio of colloquial to literary style)
  • Excessive particles: Whether particles are excessively clustered in a single sentence
  • Excessive modifiers: Whether unnecessary "this/that/those" are repeated, especially "that" with no referent
  • Spelling and spacing: Notation that violates Korean spelling rules (saisiot, tense consonants, spacing of auxiliary verbs, etc.)
  • Grammar: Subject-predicate agreement, particle misuse, confusion of passive/causative, tense inconsistency
  • Word choice: Words that do not fit the context, confusion of similar words (~로서/~로써, 맞추다/맞히다)
  • Content consistency: Whether there is information contradicting the surrounding context
  • Clichés/stereotypes: Whether idiomatic expressions are used without intent

Step 3: Ask the author

For each sentence that needs changing, propose in this format:

**문장 N**: "원문 문장"
→ 제안: "수정 문장"
- 근거: [구체적 기준, 예: "작품 전체가 평어인데 이 문장만 경칭(~하셨다)", "조사 '을/를'이 한 문장에 3회 반복"]
- 유형: [경칭/평어 | 조사 정리 | 관형어 정리 | 문체 통일 | 맞춤법 | 띄어쓰기 | 문법 | 어휘 선택 | 내용 수정]

Sentences with no issuesare explicitly marked as "no change." Do not skip them — the author must be able to confirm that every sentence has been reviewed.

When there are multiple options (e.g., when different expressions exist per version), list the options, present the rationale for each, and leave the choice to the author.

Step 4: Confirm & next paragraph

When the author approves (or gives revision instructions), reflect it in the confirmed text and move on to the next paragraph.

Precautions when proposing

"It's more natural" is not a rationale. Always specify concretely which criterion the judgment is based on. If you cannot provide a rationale, do not propose it.

The original may be better. If there is no issue in light of the criteria, do not bother fixing it. Not every sentence has to be changed.

Distinguish dialogue from narration. Non-standard notation in dialogue (dialect, colloquial style) is highly likely intentional. Treat it differently from non-standard notation in narration.

Be cautious with spelling suggestions. Point out only certain errors. For debatable notation (e.g., spacing of some compound words), mention it only to the extent of "some hold this view" and leave the judgment to the author.

Progress state management

Since the work proceeds across long texts:

  • Accumulate and save the confirmed text (as a file during the work).
  • Display the progress rate at the start of each paragraph.
  • Keep the confirmed text so that work can be resumed even if interrupted.

Evidence base: four-layer analysis

It is an analysis tool for explaining "why change it this way" in the revision loop. It can also be used as a standalone evaluation report, but its main purpose is to provide grounds for revision.

Theme

  • Infer the work's core intent in one sentence and confirm it with the author
  • Measure the frequency and placement of key vocabulary/imagery that hints at the theme
  • Identify scenes that appear unrelated to the theme (deviation rate)
  • Judge whether the ending completes/leaves open/contradicts the theme

Content

  • Points where a character's actions or speech style conflict with the setup
  • Factual conflicts in timeline, proper nouns, and spatial arrangement
  • Order of information introduction (whether information the reader needs is presented at the right time)
  • Sensory density (scenes with only narration and no sensory detail)
  • Showing vs telling ratio

Structure

  • Narrative function tag per scene (introduction/development/turn/crisis/climax/resolution)
  • Discrepancy between per-scene tension flow and the intended curve
  • Balance between per-scene length and narrative importance
  • List of point-of-view and tense shift points

Sentence

  • Sentence-ending patterns: ratio and distribution of colloquial style (~지, ~더랬지, ~거든) vs literary style (~었다, ~였다)
  • Honorific/plain speech patterns: level of deference by character and narrator
  • Particle and modifier frequency: measured per sentence
  • Sentence length distribution: rhythm analysis
  • List of spelling, grammar, and vocabulary errors

Use of quantitative analysis

Items that can be measured are automated with scripts:

  • Character count/sentence count (per scene)
  • Sentence-ending pattern ratios (regular expressions)
  • Detection of honorific/plain speech expressions
  • Measurement of particle frequency
  • Frequency and position mapping of key vocabulary
  • Sentence length histogram

With this kind of data, instead of saying "the style wavers," you can say "from paragraph 6 onward, the colloquial style ratio drops sharply from 15% to 3%."


Version comparison mode

When multiple versions of a manuscript are given:

  1. Extract the text of each version (if HWPX, use npx hwpxjs txt ).
  2. Compare programmatically: character count, sentence-ending pattern ratios, key vocabulary frequency, structural differences.
  3. Judge which version is more faithful to the intent across the four layers and present the grounds.
  4. Organize elements common to all versions into a separate list — because these are the author's deliberate choices, they are not targets for deletion in the revision loop.
  5. After selecting a base version together with the author, enter the revision loop.

Final output

When revision of all paragraphs is complete:

  1. Organize the finalized full text.
  2. Output as an HWPX file. When generating HWPX, follow the guidance of the hwpx skill.
  3. Paragraph formatting (indentation, line spacing, font, etc.) follows the formatting of the original file.