Use these prompts as templates. Replace each bracketed field. Remove instructions that do not apply to the project. Keep inputs and outputs with spoilers in the private workspace.
Models can ignore format rules or damage markup. Before a merge, validate model output with a script.
0. Context-free project bootstrap
Use this prompt to give a new visual novel to an agent without previous chat context.
Act as the project lead for machine-assisted visual novel localization. Before target changes, read the workflow, project blueprint, round-trip/build guide, and editorial/runtime QA guide. Also read the supplied project manifest and each existing authority file.
First, do a read-only readiness audit. Identify the actual runtime source, engine and build variants, and all visible text surfaces. Identify the stable-ID and schema design, extraction/import path, chapter dependency order, and editorial reading order. Identify narrator/viewpoint spans, identity/pronoun/voice authority files, and repeated text. Identify presentation limits, current review state, and release requirements.
Verify hashes. Run baseline validators. Keep verified facts separate from assumptions.
Then return:
1. canonical artifacts and conflict-priority order
2. missing evidence and risks
3. a phase plan from canary round trip through zero-change release QA
4. the manifests, reports, and validators that must exist
5. the exact next safe action and files it may modify.
Before bulk translation, verify all these conditions:
- The running game displays a controlled canary test string rebuilt through stable IDs.
- The inventory includes all visible text.
- The story/context authority files contain the initial required information.
- A clean rebuild is reproducible.
Do not use chat history, a compiled output, or the current English draft as the only authority.
[WORKFLOW AND GUIDE PAGES]
[PROJECT FILE TREE]
[PROJECT MANIFEST OR EMPTY TEMPLATE]
[EXISTING AUTHORITIES AND REPORTS]1. Script discovery and risk scan
Use this prompt on representative excerpts from across the VN. Do not supply the complete script in one request.
Analyze the supplied visual novel source rows before translation. Do not translate the full excerpt.
Return four sections:
1. TERM CANDIDATES
A table with: source form, reading evidence, proposed English, category, aliases, line IDs, confidence, and why consistency matters.
2. IDENTITY AND SPEAKER RISKS
Possible aliases, inherited titles, time/costume variants, production-only speaker suffixes, ambiguous same-name characters, and reveal-sensitive naming. Identify two identities as the same only when the excerpt supports that conclusion. Label hypotheses.
3. RECURRING GRAMMAR RISKS
Constructions likely to produce incorrect or unnatural MT in this prose. For each construction, give line IDs, source structure, probable failure, and the required translation decision.
4. STYLE OBSERVATIONS
Narration habits, register shifts, character-specific speech patterns, fragment/list rhythm, punctuation, and domain terminology.
Distinguish observed evidence from inference. Quote only the shortest phrase needed to identify a pattern.
[PROJECT BRIEF]
[SOURCE ROWS WITH STABLE IDS AND CONTEXT]2. Terminology-authority proposal
Use the evidence below for terminology research. Do not do bulk translation. Propose additions or changes to the project's terminology authority.
For each entry return:
- canonical source form
- required English form
- source aliases/readings
- category
- status: locked, working, or review
- usage rule
- evidence line IDs
- conflicts with existing entries.
Rules:
- Give explicit in-game Latin text and ruby/readings priority over guessed kanji readings.
- Do not merge aliases or people without evidence.
- Separate neutral and derogatory terms.
- Separate organization, facility, product, rank, and person names even when they share a word.
- Identify sprite/costume/age labels that must not appear in normal dialogue names.
- Mark uncertain romanization, segmentation, or diacritics as review.
- Do not silently replace an approved term. Explain any proposed change.
[CURRENT TERMINOLOGY AUTHORITY]
[EVIDENCE ROWS]3. VN-specific uncommon-grammar guide builder
Create or extend a grammar and restructuring guide for translators of this visual novel. Limit the guide to this work.
Group repeated problems by construction. Each entry must contain:
- a concise searchable heading
- stable line ID
- a short source excerpt or abstracted source shape
- head noun and omitted subject when relevant
- a realistic bad English output from a literal translator or model
- a corrected good English output that demonstrates the required decision
- the failure type: meaning, agency, scope, certainty, reveal, register, syntax, markup, or another precise category
- a short explanation of the bad output error and why the good output preserves the source meaning
- notes on register, uncertainty, reveal timing, tags, or line breaks.
Include both `Bad English` and `Good English` in each worked entry. Show the failure directly. Use realistic bad output. Give priority to repeated patterns, reversed logical scope, incorrect actors, lost uncertainty, and unreadable English. Do not add an entry only because a sentence contains unusual vocabulary.
If context is insufficient, identify the necessary surrounding rows. Do not use later scenes to remove an earlier ambiguity.
[PROJECT BRIEF]
[CURRENT GRAMMAR GUIDE]
[SOURCE ROWS WITH BEFORE/AFTER CONTEXT]4. Reading-order, identity, voice, and wordplay audit
Before bulk translation, run this prompt. When later scenes add evidence, extend the audit.
Read the supplied visual-novel script in player order. Create four short project authority files supported by evidence. Do not translate the full script.
1. READING AND REVEAL ORDER
List chapter/route order, unlock prerequisites, narrator/POV, time period, and what the reader currently knows. Flag extraction order that differs from play order.
2. IDENTITY AND SPEECH AXES
For each recurring character, record these facts separately:
- identity/gender and English pronouns
- Japanese first- and second-person forms
- gendered or nonstandard speech markers
- public/private code-switching
- changes through the story
Do not infer pronouns from feminine or masculine speech alone.
3. REPRODUCIBLE VOICE
Describe observable English features:
- contraction rate, clause length, directness, and hedging
- question style, address terms, and vocabulary
- fragments, repetitions, and emotional displacement
- the details that each narrator notices
Personality adjectives alone are insufficient. State which features remain stable and which develop.
4. WRITING-SYSTEM RISKS
Find ruby/furigana mismatches, kanji readings/components, shared glyphs across identities, and homophones. Find script switches, name formation, visual contrasts, and recurring lexical networks. For each case, give line IDs, function, and recommendation. Select one strategy: preserve directly, explain once, rebuild locally, accept controlled loss, or do not force. Keep explicit wordplay separate from speculation.
Label observation and inference. Respect reveal timing. State what context is still missing.
[PROJECT BRIEF]
[CHAPTER/UNLOCK DATA]
[SOURCE ROWS IN PLAYER ORDER]5. Strict batch translation
Where possible, protect complex tags with placeholders before you use this prompt.
Translate only the rows in TARGET ROWS. Do not change CONTEXT ROWS. Do not include CONTEXT ROWS in the output.
Return TSV only, with exactly these columns and no Markdown fence:
line_id<TAB>target_text<TAB>notes
Requirements:
- Output exactly one row for every target line_id, in the same order.
- Copy line_id exactly. Do not add, omit, merge, or split rows.
- Do not translate speaker IDs, commands, filenames, variables, or tag syntax.
- Preserve every [[PLACEHOLDER_...]] token exactly once and in a grammatically appropriate position.
- Use approved terminology exactly.
- Resolve omitted subjects from the supplied scene context. If ambiguity remains, preserve it. Add a note.
- Preserve the source's certainty level: fact, observation, inference, hearsay, possibility, or condition.
- Preserve identity reveal timing and the name used by the current viewpoint.
- Where necessary for clear English, restructure Japanese syntax. Do not invent or omit information.
- Match each speaker's approved voice and register.
- Keep text within [LIMIT] unless a note explains why that is impossible.
- Leave notes empty unless there is a real ambiguity, terminology conflict, cultural issue, or layout risk.
[PROJECT BRIEF]
[RELEVANT TERMINOLOGY ENTRIES]
[RELEVANT READING-ORDER, IDENTITY, VOICE, AND WORDPLAY ENTRIES]
[RELEVANT GRAMMAR-GUIDE SECTIONS]
[SCENE/SPEAKER SUMMARY]
[CONTEXT ROWS]
[TARGET ROWS]6. Bilingual accuracy and continuity review
This pass identifies errors first. It does not rewrite each line to match the reviewer's preferred style.
Audit each source/target pair as a bilingual visual-novel editor.
Check, in order:
1. meaning, omitted information, and invented information
2. narrator/POV, subject, object, pronoun number, tense, and causal direction
3. negation scope, conditionals, concession, and evidentiality
4. terminology, aliases, titles, and reveal timing
5. voice, register, fluency, and textbox constraints
6. tags, placeholders, variables, and punctuation.
Return a table with:
line_id | verdict | severity | issue | minimal_revision | rationale
Verdicts: pass, revise, or needs-context.
Severity: critical, major, minor, or none.
Use the smallest revision that fixes the issue. Do not report a faithful line only because another wording is possible. If the source is ambiguous, request the specific adjacent context needed. Do not guess.
[PROJECT BRIEF]
[TERMINOLOGY AUTHORITY]
[READING/REVEAL, IDENTITY, VOICE, AND WORDPLAY AUTHORITIES]
[RELEVANT GRAMMAR GUIDE]
[SOURCE/TARGET ROWS WITH CONTEXT]7. Character voice and prose pass
Run this pass only after bilingual accuracy review. First, read the source and target together. Then read the complete target-language chapter alone.
Edit this reviewed visual-novel chapter for character voice and natural target-language prose. Treat the current target as an editable draft. Preserve all established meaning and safeguards.
For each line, consult the local exchange, speaker/narrator record, and route progression. Check:
- diction, contraction rate, clause length, directness, hedging, and question style
- terms of address, public/private code-switching, and emotional pressure
- dialogue timing, meaningful fragments/repetitions, and exposition that sounds translated
- narrator-specific noticing habits and literary texture
- idioms or jokes that need functional rather than literal equivalence.
Do not undo approved names, pronouns, reveal boundaries, uncertainty, wordplay decisions, tags, or IDs. Do not invent a character accent or verbal habit. After edits against the source, read the complete English chapter. Correct only actual continuity or readability problems.
Return:
1. changed rows as line_id | revised_target | reason
2. unresolved questions
3. a chapter sign-off stating rows reviewed, safeguards checked, and validation still required.
[PROJECT BRIEF]
[IDENTITY AND VOICE AUTHORITIES]
[WORDPLAY/GRAMMAR AUTHORITIES]
[SOURCE/TARGET CHAPTER IN PLAYER ORDER]8. Consistency and reveal audit
Audit these translated batches against the terminology authority and timeline notes.
Report only actionable findings in these groups:
- non-approved spelling or capitalization
- deprecated term
- alias used in the wrong social or timeline context
- production-only speaker qualifier leaked into displayed text
- two distinct concepts collapsed into one English term
- same source concept translated inconsistently without a contextual reason
- known identity named before the viewpoint/source reveals it
- title, kinship term, or rank flattened in a way that changes the relationship.
For each finding, give line_id, current text, required form or question, authority entry, and confidence. When grammar or context changes the correct form, check each occurrence before replacement.
[TERMINOLOGY AND IDENTITY AUTHORITIES]
[READING-ORDER/TIMELINE NOTES]
[VOICE AND WORDPLAY LEXICAL-NETWORK RULES]
[TRANSLATED ROWS]9. Technical and row-correspondence QA
After accuracy and prose edits, run this mechanical release check. Do not use it for style changes. If the check finds defects, apply corrections. Validate the corrections. Repeat the full-corpus check until the result is CLEAN.
Perform a complete final QA of these source/target rows. Do not propose optional rewrites.
Verify every row for:
- exactly one matching stable ID, correct order, and English that corresponds to that Japanese row
- no missing, duplicated, untranslated, truncated, merged, or displaced text
- spelling, grammar, punctuation, quotation pairing, capitalization, and locked romanization/terminology
- exact preservation of tags, variables, placeholders, escaped newlines, and control syntax
- correct speaker/narrator attribution and first-person singular/plural forms in context
- allowed versus accidental source-language remnants.
Return only actionable findings as:
line_id | current_target | required_target | category | evidence
If and only if every supplied row passes, return exactly `CLEAN`. Inspect the current files. A previous clean report does not verify the current files.
[PROJECT BRIEF]
[AUTHORITY FILE INDEX IN CONFLICT-PRIORITY ORDER]
[SOURCE/TARGET ROWS IN PLAYER ORDER]
[CURRENT VALIDATOR OUTPUT]10. Handoff and completion audit
Before project handoff or phase completion, run this audit.
Audit the project state. Do not translate or edit lines.
Verify and report:
- canonical source/version/hash and rejected alternate sources
- exact source, target, chapter, and support/UI coverage counts
- unique stable IDs, target/archive equality, and review statuses
- tag, variable, placeholder, and control-field integrity
- remaining source-language text, with explicit allowed exceptions
- unresolved names, terminology, identity, pronoun, reveal, wordplay, and layout issues
- completed review gates per chapter in player order
- changelog completeness and validator commands/results
- compile/import state and in-engine QA state.
Return the following:
- COMPLETE, INCOMPLETE, or BLOCKED
- actionable findings
- authoritative files in conflict-priority order
- exact next action
- a short handoff brief
Nonblank targets do not prove that a first draft has passed review.
[PROJECT BRIEF]
[PROJECT MANIFESTS AND REPORTS]
[VALIDATOR OUTPUT]11. Repeated and parallel passage reconciliation
Audit the full bilingual corpus for identical and near-identical Japanese passages. Include quotations, documents, flashbacks, retellings, alternate viewpoints, repeated scenes, ritual phrases, and key concepts. Normalize markup only to find candidates. Keep stable IDs and full scene context for decisions.
For each cluster, compare source wording, speaker/narrator, time, reveal state, dramatic function, tags, and existing English. Use the same English when the source and function are the same. If context requires a difference, preserve it. Record the exact reason. Report accidental variation in locked terminology, capitalization, titles, names, and repeated images. Report duplicate adjacent English words or linked terms.
Return stable-ID patch proposals and a cluster report. Do not apply a global replacement without checking every context.
[PROJECT BRIEF]
[REPEATED-SOURCE CANDIDATE REPORT]
[BILINGUAL ROWS WITH SCENE CONTEXT]12. Support, presentation, and stateful runtime QA
Audit the built game and the scenario table. Include choices, speaker labels, tips/glossary, menus, settings, popups, galleries, credits, and chapter select. Include textures, ruby/helper text, backlog, fonts, wrapping, page breaks, tags, portraits, backgrounds, and voice timing. Test each supported input method and resolution.
Test these stateful sequences:
- Repeatedly open and close tips.
- Click nested or multiple linked terms.
- Return to dialogue.
- Advance with keyboard, mouse, controller, wheel, auto, and skip.
- Open backlog, settings, save/load, choices, and chapter select.
- Move focus away from the game, then back to it.
- Resize the window.
- Switch full-screen mode.
- If supported, switch language or asset packs.
Confirm that overlays release input focus. Confirm that words never split at apostrophes. Confirm that portraits never cover text. Confirm that each page is accessible.
Record the exact environment, starting state, actions, and expected and actual results. Include a screenshot/log/line ID, affected scope, and regression case. After a correction, search for other instances with the same structure. Repeat the full test case. Do not check only the original screenshot.
[PROJECT BRIEF]
[QA MATRIX]
[BUILT GAME AND DEBUG INSTRUCTIONS]
[KNOWN FAILURES]13. Clean build, patcher, and release audit
Evaluate release readiness from an immutable supported installation. Verify exact input hashes. Rebuild with only canonical reviewed targets. Compare the build report and output hashes.
Test the following:
- fresh installation and previous-version update
- reinstallation and repeated installation
- recovery from an interrupted state
- removal/restoration and save compatibility
- paths with spaces or non-ASCII characters
- each supported executable/store/platform variant
Before changes, the patcher must detect unsupported, fresh, already-patched, and partial states. It must complete safely or leave the installation recoverable. If the patcher changes language packs or restarts the game, verify the visible state after restart. File checks alone are insufficient.
Return the following:
- COMPLETE, INCOMPLETE, or BLOCKED
- exact tested matrix
- failures with reproduction steps
- files changed
- rollback result
- release hashes
- the one command that reproduces the distributable
Do not approve a release only because it works in the developer's existing installation.
[PROJECT MANIFEST]
[CLEAN-BASE AND RELEASE HASHES]
[BUILD REPORT]
[QA MATRIX AND PATCHER LOGS]Prompt-version record
Record a small manifest beside generated batches:
prompt_version: [git commit or semantic version]
model: [exact model identifier]
generated_at: [ISO 8601 timestamp]
source_batch_hash: [SHA-256]
terminology_version: [git commit/hash]
grammar_guide_version: [git commit/hash]
reading_order_version: [git commit/hash]
identity_voice_version: [git commit/hash]
wordplay_version: [git commit/hash]
settings: [temperature, seed, or provider options]
human_reviewer: [name or blank]A future model call can produce different prose. Reproducibility requires records of the exact input, rules, output, and review history.