公開用翻訳

# Publication Translation

Translate and refine text with <integration-tag>direct:llm_inference</integration-tag> while preserving facts, terminology, and Markdown structure.

## Choose a mode

- `quick`: one-pass translation for short, low-risk text.
- `normal`: analyze domain, audience, voice, and terminology before translating.
- `refined`: analyze, draft, critique, revise, and polish for publication.

Confirm source language, target language, audience, style, and required glossary. If the user did not choose a mode, use `normal`. For long input, split at Markdown block boundaries and keep a shared glossary across chunks.

## Prompt contract

Inspect the live Action schema. Put stable translation principles in `systemPrompt` and the source plus mode-specific task in `prompt`. Never place secrets, private credentials, or unrelated context in either field.

Require the model to:

- preserve facts, numbers, links, headings, lists, tables, code, and emphasis;
- use idiomatic target-language phrasing without adding unsupported claims;
- apply the approved glossary consistently;
- keep proper nouns stable unless an established translation exists;
- mark genuinely ambiguous passages instead of silently resolving them.

## Review

For `normal`, check omissions, terminology, tone, and formatting before delivery. For `refined`, use separate calls for draft, critique, revision, and final polish so each stage has a clear input. Show the user material interpretation choices and keep the original text available for comparison.

The Action returns plain text. Do not claim automatic file creation, URL fetching, persistent preferences, parallel subagents, or lossless translation. Ask the user to provide text that fits the live input limits.

## Attribution

This native Loadout re-authoring condenses the source package's quick, normal, and refined methods and binds each generation stage to Aident-managed LLM inference. See [UPSTREAM.md](UPSTREAM.md) and [LICENSE.txt](LICENSE.txt).