3.1 Prompt Design and Practical Refinement¶
What you will learn on this page
- Core principles of prompt design (5W1H, role assignment, few-shot prompting)
- Prompt strategies for refining existing English
- A “fix your claim in Japanese first, then convert to English” approach
- Step-by-step prompt refinement (iterative improvement)
- A template library by purpose (style, proofreading, tense checks, logical flow, comprehensive checks)
- How to compare and evaluate AI outputs
- When to stop refining (a practical stopping rule)
- How to leverage saved custom instructions
How this page fits in
On 2.3 Academic Style, Tense, and Logical Flow, we explained rules of academic style, tense conventions, logical flow, and pre-editing for Japanese drafts intended for translation. This page builds on that knowledge and focuses on a practical workflow for designing prompts and refining English step by step.
Core principle: Do not let AI produce the “answer”¶
Generative AI can be used either as a tool that helps you develop your English or as a tool that replaces your English. A key way to stay on the “development” side is to fix AI’s role as an amplifier of your thinking and expression, rather than asking it to produce the content for you.
Recall the Augmented Competence model introduced in 1.2 Academic Integrity and Plagiarism Prevention. Refinement with AI assumes you stay within your own receptive competence. If you cannot ask yourself “Why is this better?” or “Do I fully understand this expression?”, the process shifts from augmentation to replacement.
It helps to define a minimal personal rule set to reduce mistakes.
Example personal rules
“Structure consultation, example phrases, paraphrase suggestions, grammar explanations, and checking are OK. Fabricating content, generating citations without evidence, and producing a full submission as-is are not allowed.”
Principles of prompt design¶
The basics¶
Output quality changes dramatically depending on the prompt. When people feel that “AI is not useful for academic writing,” the prompt is often vague, or the user stops after one attempt.
Three principles:
- Provide more detailed context: Instead of “Fix this,” specify criteria, perspective, and constraints.
- Give clear and precise instructions: AI follows instructions literally, so reduce ambiguity.
- Iterate: Outputs vary. Keep the conversation going by stating what you dislike and what you want improved.
Yanase (2023) also notes two key cautions for style improvement with ChatGPT.
Caution 1: Give objective, explicit instructions
Treat prompting like programming. State the goal, audience, context, and constraints explicitly. Even for translation, specifying the purpose and target audience improves output quality (Yamada, 2023).
Caution 2: Do not assume human-like understanding
Better prompts improve responses, but AI does not “understand meaning” like humans. If you write “Improve the flow,” it may emphasize aspects you did not intend. Make instructions more concrete, for example by adding “The core of the argument is …” or “by emphasizing …”.
Practical techniques¶
- Assign a role: e.g., “You are an expert academic editor.”
- Use delimiters and structure markers:
#for sections,<<< >>>for quoted input, etc. - Use few-shot prompting: Provide input-output examples when possible. If you cannot provide examples, detailed constraints become more important.
5W1H for prompts
Including the elements below improves stability.
| Element | What it means | Example |
|---|---|---|
| What | Target to improve | “The Methods section below” |
| Why | Purpose | “To make it natural academic English” |
| How | Criteria | “Short subjects, passive voice, past tense” |
| How many | Output count | “Provide three options” |
| Constraints | What not to do | “Do not change meaning; do not add content” |
Common prompt failures and fixes¶
| Failure pattern | Why it fails | Better instruction |
|---|---|---|
| “Please fix this English.” | Too vague, triggers a full rewrite | “Only point out grammar and usage errors. No style changes.” |
| “Make it perfect English.” | “Perfect” is undefined; may become overly formal | “Make it natural for readers in [your field].” |
| “Translate this into English.” | Can become literal; ignores section conventions | “Write this as part of a Methods section in past tense, mostly passive.” |
| Too many constraints (10+) | AI drops constraints | Limit to 5 or fewer, put the most important first |
Fix the claim in L1 first, then convert to English¶
When you struggle in English, it becomes easy to be pulled into AI-generated phrasing. A safer approach is to first stabilize what you want to say in your first language (L1). This is a practical version of the pre-editing approach in 2.3.
Steps:
- Write one conclusion sentence and two reasons in Japanese.
- Ask AI to convert it into English.
- Ask AI to check for deviation from the Japanese original.
Prompt example (Step 2):
Convert the following Japanese claim into natural English at around CEFR B2.
Use short sentences with clear subjects and verbs.
Do not add any new content.
[Japanese text]
Prompt example (Step 3):
Point out any parts of the English below that deviate from the original Japanese.
List the deviations in bullet points.
Japanese: [Japanese text]
English: [Generated English]
This keeps the division clear: content is yours; language support is AI’s.
Specify English level and refine step by step¶
Instead of expecting one prompt to produce a final version, use a staged process.
A three-stage refinement process¶
| Stage | Goal | Main focus |
|---|---|---|
| Stage 1: Convert content into English | Get correct meaning | Accuracy, completeness |
| Stage 2: Adjust to academic style | Fit section conventions | subject length, hedging, tense, active/passive |
| Stage 3: Polish expressions | Improve naturalness | collocations, word choice, rhythm |
Do not ask for everything at once
If you combine too many requirements, output becomes unstable. In one prompt, focus on one or two improvements.
Example: staged refinement (Stage 1 → Stage 2)
What to do in Stage 3¶
In Stage 3, you should validate expressions against your field’s papers and corpora (see 2.2 Checking Conventional Expressions with Corpora). Use Google Scholar phrase searches and tools like AntConc to confirm that AI-suggested expressions are actually used in your discourse community.
Turn vocabulary and collocations into “your own list”¶
A useful approach is to move from “AI gives phrases” to “I reuse phrases intentionally.”
Suggest three places in the paragraph below where I can replace wording with widely used academic phrases.
For each suggestion, add:
(1) when it is appropriate to use, and
(2) one caution.
[Paragraph]
Then select only the phrases you plan to reuse, add them to a personal phrase list, and set a rule such as “Use at least two phrases from my list in the next writing session.”
Prompt: build a section-based personal phrase list
From the English below, extract reusable academic phrases and classify them by section:
- Introduction (background, gap, purpose)
- Methods (procedures, analysis)
- Results (reporting, figure/table reference)
- Discussion (interpretation, comparison, limitations)
For each category, provide 3–5 phrases and a one-line usage note.
[Your draft]
A practical mixed-language drafting workflow
One approach that works well for me in practice is:
-
Draft quickly, mixing Japanese and English as needed (use familiar field-standard terms and phrases).
-
Use AI to first make the English correct: I am writing an academic paper. Please translate the following sentences written in Japanese into natural English.
-
Then ask for a stronger refinement: Proofread this, significantly improving clarity and flow.
Because AI sees the full context, this can outperform pure machine translation, especially for technical academic writing.
Template library by purpose¶
Saving templates you use often improves efficiency. The templates below are general-purpose and are designed to leverage the style, tense, and logic guidance in 2.3.
Core templates by section¶
Template 1: Japanese (L1) → English conversion¶
Convert the following Japanese into natural academic English as a [SECTION NAME] section.
Constraints:
- Tense: [past / present / follow section conventions]
- Voice: [mostly passive / mostly active / mixed]
- Hedging: [add appropriately / do not add]
- Do not add new content
- Do not generate or invent citations
[Japanese text]
Template 2: Diagnose an English draft¶
Diagnose the following English as a [SECTION NAME] section.
For each item, write either OK or point out a concrete issue:
(1) Grammar and usage
(2) Tense conventions for the section
(3) Academic style (subject length, hedging, wordiness)
(4) Logical flow (sentence links)
For each issue, provide at most one suggested fix.
[English text]
Template 3: Generate paraphrase options¶
Paraphrase the TARGET below without changing meaning.
Provide 3 options. For each option, include:
1) best use case (one line)
2) formality level (high/medium/low)
TARGET:
<<<
[target expression]
>>>
CONTEXT:
<<<
[English with surrounding sentences]
>>>
Style improvement¶
Prompt: general academic style improvement
Prompt: Japanese (L1) → Introduction paragraph
Methods templates
Methods-specific templates (converting bullet points, participant description templates, removing vague wording, statistical reporting formats) are compiled in: 3.3 Writing the Methods Section
Prompt: detect and reduce wordiness
Proofreading¶
The prompts below are useful for proofreading, depending on how strongly you want AI to intervene.
Prompt: adjustable proofreading strength
A two-step approach often works well: broad proofreading first, then targeted constraints.
Step 1 (broad):
Stronger intervention:
Light intervention:
Minimal intervention (typos/grammar only):
Step 2 (add targeted checks):
Prompt: revise only where definitely necessary
Proofread the following sentences and provide both the
original sentences and your suggested revisions. If they are
acceptable, you don't have to make suggestions. However,
when revisions are definitely necessary, please clearly indicate
what the original sentence was and how you recommend it
should be revised.
Prompt: shorten while preserving content
# Task
Suppose you're a professional proofreader. Your client asks you to
shorten the sentences by following these rules.
# Rules
1. Preserve the original wording as much as possible without
sacrificing clarity or brevity.
2. Do not skip contained information.
3. Keep the quotes as they are.
4. Mark the changed parts so that they can be easily identified.
Tense and logical flow checks¶
Prompt: reporting verb tense check
Check the tense of reporting verbs used to cite prior research in the Introduction below.
Criteria:
- Specific author + year citations → past tense
- Summaries of multiple studies → present perfect
- Widely established knowledge → present tense
If there are problems, explain why and provide one fix per case.
[Introduction text]
Prompt: tense and logic check (paragraph)
Check the paragraph below for:
(1) tense consistency appropriate to the section (Methods/Results/Discussion)
(2) logical flow (gaps between sentences, overuse/underuse of connectors)
If there are problems, explain the fix strategy and provide one revised version only.
Section: [section name]
[Paragraph]
Prompt: causal language check
Information structure and paragraph structure checks¶
Prompt: given-new and demonstratives
Prompt: paragraph structure check
For the paragraphs below, check:
(1) whether each paragraph has a clear topic sentence,
(2) whether each paragraph sticks to one topic,
(3) whether the sequence of topic sentences alone forms a coherent argument.
Identify problematic paragraphs and suggest improvement directions.
Do not rewrite.
[Text]
Prompt: detect logical leaps
Comprehensive checks¶
Prompt: style consistency
Check the manuscript below for style consistency.
Items:
(1) First-person use: consistent (we vs I), and active/passive switching is systematic
(2) Hedging: similar claim strength uses similar hedging strength
Point out inconsistencies and recommend a consistent choice.
[Manuscript]
Mechanical checks for target venue style (spelling variant, numbers, abbreviation rules) are summarized in:
4.1 Grammar and Style Checks
Prompt: comprehensive paragraph-quality check
Check the paragraph below across five dimensions.
For each, write OK or point out a concrete issue.
(1) Tense: appropriate and consistent for the section
(2) Information structure: follows given-new
(3) Logical flow: no leaps between sentences
(4) Connectors: appropriate and appropriately formal
(5) Paragraph structure: clear topic sentence, one topic
Section: [section name]
[Paragraph]
Prompt: paraphrase practice
Comparing and evaluating AI outputs¶
Even with the same prompt, different models or runs can produce different outputs. Learning how to compare outputs helps stabilize quality.
Evaluation dimensions¶
| Dimension | What to check | Practical criterion |
|---|---|---|
| Accuracy | no deviation from your intended meaning | compare with your Japanese original |
| Naturalness | fits academic English in your field | compare with papers in your field |
| Specificity | not overly generic | includes your study-specific details |
| Consistency | fits surrounding context | read with other sections |
| Conciseness | avoids redundancy | can you shorten without loss? |
Prompt: compare two AI outputs
The two English versions below were generated from the same Japanese.
Compare them on five dimensions and decide which is better for each.
Dimensions:
(1) accuracy, (2) academic naturalness, (3) specificity,
(4) contextual consistency, (5) conciseness
Japanese: [Japanese]
A: [Output A]
B: [Output B]
Using multiple AI models¶
Different models have different tendencies. It can be effective to choose models by purpose.
- Learn tendencies by running the same prompt across models (ChatGPT, Claude, Gemini).
- Allocate tasks by strengths (e.g., grammar checks vs paraphrase options).
- Make the final choice yourself based on your field and your style.
See also: 4.1 Grammar and Style Checks.
When to stop refining¶
Refinement can be repeated endlessly, but over-refinement has costs:
- Loss of authorial voice (the text becomes “generic”)
- Drifting beyond Augmented Competence (you adopt expressions you cannot justify)
- Diminishing returns (small gains after several rounds)
Practical stopping criteria¶
When the following are true, it is usually time to stop:
- Meaning matches the Japanese original (no deviation)
- Tense and voice match section conventions
- You can explain every expression you use
- Style is consistent with the surrounding sections
- Feedback begins to feel like preference rather than improvement
If you still feel “it must get better” after five rounds
You may be adapting your judgment to AI output. Pause, and reread your own draft from the start.
Using saved custom instructions¶
Many AI tools allow you to save instructions so you do not have to repeat them each time. This is useful for repeated proofreading or for sharing common rules within a lab.
| Tool | Feature name | Creating custom instructions | Attach references | Notes |
|---|---|---|---|---|
| ChatGPT | GPTs | often requires paid plan | Yes | Some public GPTs available broadly |
| Claude | Projects | often requires paid plan | Yes | |
| Gemini | Gems | often available | Yes |
If you store your target venue’s style guide and your own rules, you can dramatically shorten prompt setup.
Plans may change
Pricing and availability change frequently. Check official pages for the latest information.
Use a prompt-sharing site
Prompt Me-Mo lets you save prompts and share them. You can use it without signing in. Register the templates you use often so you can reuse them instantly while writing.
