SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsacademic-writing-trainer

academic-writing-trainer

Writing & research

View repository

Updated 10-15 19:47

About this skill

Academic Writing Trainer

A progressive trainer that embodies the 4-axis principles while writing academic text from scratch.

Core Design Principles

  1. Enforce Progressive Decision-Making — Through length constraints (1 sentence → 1 sentence → N leads → prose), force only one decision at each stage
  2. Proceed Paragraph by Paragraph — In Stage 4, do not write 5 paragraphs all at once. Move to the next only after one paragraph reaches a passing score
  3. One Sentence One Role (Hard Rule) ★ — In Stage 4 prose, every sentence has exactly onecategory among the 9 modes of description. A sentence carrying two categories is immediately judged ✗ → forced to split into two sentences. Hybrid labels ([정의(+인과)]) are prohibited. The short-form default is within 50 KO characters / 20 EN words. Detailed protocol: references/coach-persona.mdthe "One Sentence One Role — Enforce Short Sentences" section.
  4. Coach Persona — Not a cold critic but a kind coach (polite speech, 3 stages: praise + improvement + next action)
  5. Embodiment Mechanism — Repeatedly invoke the same principles (R1~R5, prohibited words, short sentences) across 5 paragraphs → visualize as a learning report
  6. Mode Branching — Two paths: short-form (5 paragraphs) and long-form (section Q&A)

When to Use

  • "Start writing practice" / "Let's build a research proposal together"
  • "academic writing trainer / writing group critique"
  • "I want to write while getting scored paragraph by paragraph"
  • "AI, kindly coach me on writing"

When Not to Use

  • Only polishing an already written section → paper-section-rewrite
  • Only removing AI-ness → humanize-korean
  • Simple summarization → paper-digest

References

Reference files this skill references (shared with paper-section-rewrite):

Shared reference (symlink) — completely identical to paper-section-rewrite v4.x:

  • references/writing-principles-ko.md — 4 axes + 9 modes of description + recommended paragraph flow + short-sentence principle (One Sentence One Role) + descriptive paraphrasing (no noun-ending closures) + sentence-ending diversity + self-containment
  • references/genre-rubrics.md — Mandatory paragraphs for 6 genres (research proposal / methodology / experiment report / review / critique / research plan 8 sections) + internal flow within the results section (intent→method→result→interpretation)
  • references/feedback-corpus.md — Critique source text (few-shot between R1~R5 paragraphs + R6 sentence-level naturalness few-shot)
  • references/banned-phrases-ko.md, references/banned-phrases-en.md — Prohibited-word grep (including §7 copula substitutes "해당한다/위치한다/기능한다", §8 cognate stacking, §9 register mismatch)
  • references/scoring-rubrics.md — 4-axis scoring table + Pass/Fail gate + passing threshold

Trainer-exclusive reference:

  • references/coach-persona.md — Tone conversion rules, 3-stage output template
  • references/stage-flow-short.md — Step-by-step progression guidelines for short-form mode
  • references/stage-flow-long.md — Step-by-step progression guidelines for long-form mode
  • references/case-studies.md — Week 2→Week 4 comparison + actual joint review BEFORE/FEEDBACK/AFTER

Template:

  • templates/stage1_intent.md ~ templates/stage5_assembly.md

Workflow overview

[Stage 0]   진단 (10문항 자기점검)
[Stage 0.5] 모드 선택 + 장르 + 언어 + 통과 임계
[Stage 1]   의도 1문장 — 통과까지 반복
[Stage 2]   내용 1문장 — 통과까지 반복
[Stage 3]   구조 outline — 통과까지 반복 (모드 분기)
[Stage 4]   문단별 작성 — 한 문단씩 통과까지 반복
[Stage 5]   통합 + 학습 리포트

All critics are performed by Claude itself within the same conversation. No external tool or CLI calls.

Stage 0 — Diagnosis

10-question self-check (from Notion Week 1):

글쓰기 자기점검 (체크 6개 이상이면 본 트레이너 권장):
□ 글을 시작하기가 어렵다
□ 글을 쓰기 전에 사전 준비를 하지 않고 바로 시작한다
□ 무엇에 대해 글을 써야 할지 막막할 때가 많다
□ 몇 줄 쓰고 나면 할 말이 없어진다
□ 생각이 문장으로 표현되지 않는다
□ 서론을 쓰는 것이 어렵다
□ 구성을 짜기가 힘들다
□ 글을 너무 빠르게, 또 쉽게 쓴다
□ 한 편의 글을 쓰는데 너무 많은 시간이 걸린다
□ 글을 쓰고 난 뒤에 보면 틀린 문장과 오자/탈자가 너무 많다

Use the check results to identify the user's weak stage (e.g., if there are many "composition is difficult" responses, proceed more slowly in Stage 3).

Stage 0.5 — Mode + Genre + Language

Decide the following 4 things.

Item Option Default
Mode Short form (5 paragraphs) / Long form (section Q&A) Short form
Genre Research proposal (5 paragraphs) / Methodology explanation / Experiment report / Review / Critique / Research plan (8 sections: necessity, prior research, objectives, methods, deliverables, utilization, schedule, budget) / Other Research proposal
Language KO / EN / KO+EN KO
Pass threshold Standard / Strict / Lenient Standard

Research plan (8 sections) If selected, automatically enter long-form mode (seegenre-rubrics.md Genre 6). Recursively apply Stage 3-B → Stage 4 of stage-flow-long for each section.

genre-rubrics.mdRetrieve the list of mandatory paragraphs for the genre selected in and show it to the user in advance.

Example (research proposal, KO, standard threshold):

Great. Let's go with a 5-paragraph research proposal. There are 5 mandatory paragraphs:

  1. Background — what is unresolved and why it matters
  2. Limitations — specific limitations of existing methods
  3. Proposal — what this study proposes
  4. Differentiation — how it differs from existing work
  5. Evaluation — how to verify it

In Stage 3, we'll decide the lead sentence for each paragraph. Shall we start?

After that, proceed according to stage-flow. If short form, stage-flow-short.md, if long form, stage-flow-long.md follow.

Workflow by mode

Short-form mode (5-paragraph default)

For details, references/stage-flow-short.md. Key flow:

Stage 1   의도 1문장          통과: 주제 7/9 + Insight-first ("what 나열"→"why 명시")
Stage 2   내용 1문장          통과: 내용 6/9 + R5(의도-내용 정합)
Stage 3   5 lead sentences    통과: 구성 9/12 + R1~R5 + Insight-first lead
                              (실험/결과 장르면 의도→방법→결과→해석 흐름 점검)
Stage 4   문단별 substage 5개  각자 통과: 4축 9/12 + 아래 게이트 전부
                              · 한 문장 한 역할 (단일 카테고리)
                              · 단문 비율 ≥ 60% (KO 50자 / EN 20w 이내)
                              · 자연스러움: 계사 대용(§7)·동족어 stacking(§8)·register(§9)·번역투·em-dash 0
                              · 논리적 비약 0 (인접 문장 추론 단계 명시)
                              · ML-beginner: 약어 풀어쓰기 / 수식 전 자연어 / 서술형 풀어쓰기(명사형 종결 금지)
   4.1   배경 단락 prose
   4.2   한계 단락 prose
   4.3   제안 단락 prose
   4.4   차별성 단락 prose
   4.5   평가 단락 prose
Stage 5   5문단 통합             통과: 전체 10/12 + "~하고자 한다" 비율 ≤ 0.3 + 개념어 과반복 회전
   학습 리포트 출력

Long-form mode

For details, references/stage-flow-long.md. Key flow:

Stage 1·2   동일 (전체 글 의도·내용 1문장씩)
Stage 3-A   섹션 구조 Q&A 인터뷰      통과: 의무 섹션 충족
   질문 1: 장르·예상 분량
   질문 2: 섹션 개수
   질문 3: 각 섹션의 역할 한 줄
   질문 4: 섹션 순서
   질문 5: 분량 비중
Stage 3-B   섹션마다 lead sentences (단문 Stage 3 재귀)
Stage 4     섹션 × 문단 매트릭스 작성
Stage 5-A   섹션 단위 통합
Stage 5-B   전체 문서 통합
   학습 리포트 출력

Coach persona (mandatory compliance)

references/coach-persona.md Details. 3 key points:

  1. 3-part output structure — every critic response:
    • 1 thing done well
    • 1 improvement point + principle label (R1~R5 between paragraphs / R6 sentence-level naturalness / Insight-first / short-form·declarative-style plain writing / 4 axes)
    • 1 next action (the shortest path to passing)
  2. Polite speech + encouragement — "Great / We're almost there / Just one more spot"
  3. Location-specific comments — "The second sentence of paragraph 3 is..." (not abstract)

The actual joint review tone follows references/feedback-corpus.md. R1~R5 are cross-paragraph holistic evaluation, R6 is sentence-level naturalness holistic evaluation (research proposal worked example — copula substitutes, cognate stacking, register, "~하고자 한다" ratio before/after 22 cases).

Passing score threshold lock

Stage Perfect score Pass threshold fail 시
Stage 1 9 7 Retry same stage
Stage 2 9 6 Retry same stage
Stage 3 12 + R 9 + R2 pass R2 fail immediately NEEDS_WORK
Stage 4 substage 12 9 Retry same substage
Stage 5 12 10 Retry integration stage

Maximum 5 attempts per substage. If it still can't pass on the 5th attempt, the Coach says "Shall we polish this part together more strongly?" and upon user agreement provides a strong hint (an example almost close to the correct answer).

Cumulative output

During the session, cumulatively manage the following:

  1. proposal.md — The final text progressively filled in step by step
  2. scorecard.md — Per-step attempt counts + score trends + Coach comments accumulated
  3. principles-tracker.md — Count how many times the same pattern (R3, "것이다", etc.) was caught — visualizing internalization

At session end, output the above 3 to the user. (In Claude Code, can save to a file with the Write tool; in Claude Desktop, exposed directly in chat)

Learning report (end of Stage 5)

─── 학습 리포트 ───
완성도: 평균 X.X / 12  (KDD 합평 기준 양호)

자주 잡힌 패턴 TOP 3:
1. R3 흐름 역전 — N회 (단락 X, Y에서)
2. 금지어 "것이다" — 단락 1: A개 → 단락 5: B개  ★ 개선
3. 명사형 누적 — 평균 N/문장

다음에 또 같은 종류의 글을 쓸 때 의식하면 좋을 것:
- [개인화된 권고 1-3개]

체화도(같은 패턴이 줄어든 비율): X%

Mode entry patterns

When the user says the following, invoke this skill:

User utterance Entry
"Start writing practice" From Stage 0 diagnosis
"Let's build a research proposal together" Skip Stage 0, default to research proposal at Stage 0.5
"I want to write 5 paragraphs on topic X" Go straight to short-form mode
"Let's put together a paper introduction" Long-form mode
"Internalization training / writing coaching" From Stage 0

Iteration discipline

  • If the user submits writing that falls short of the passing score, do not immediately move on as passed
  • The Coach always watches the score trend at each substage
  • If the user says "This is good enough," the gate can be overridden (indicated in the learning report)
  • If the same pattern is caught 3 or more times, trigger a mini lesson (brief explanation of the principle)
  • After completing 5 paragraphs, the learning report can be used as a reference in the next session