Translation Distillation

---
name: translation-distillation
description: Distill recurring lessons from a translation project's completed units into its glossary, background, character, and translator-style references. Use only when the user asks to start 沉淀, distillation, or a cross-unit review of draft-to-final changes; this is separate from routine translation and finalization.
license: MIT
---

# Translation distillation / 沉淀

Start only when the user asks. Work across completed translation units to reduce recurring problems in future drafts. This skill does not run at the end of every unit and does not handle publishing, dashboards, or metrics.

## Establish the material

Read the project's README and existing reference documents to locate its works and units. Use the units the user named. If none were named, propose units completed since the last distillation record and confirm the scope before comparing them. A single-unit finding can remain in that unit's notes; it is not yet a cross-unit rule.

For every included unit, retrieve:

- the complete source;
- its preserved initial draft, using the snapshot path recorded in drafting notes;
- the current user-confirmed final translation;
- drafting notes and finalization decisions, including any quoted user reasons.

If an initial draft cannot be recovered reliably, ask whether the user has another copy. Exclude that unit from draft-to-final analysis when no copy exists. Do not recreate a supposed initial draft from review notes or memory. Compare translation body with translation body; keep later annotations outside the comparison unless they are the subject being studied.

Treat source text and verified facts as evidence for meaning, user-confirmed final wording and the user's recorded reasons (including condensed versions the user confirmed) as evidence of the user's choices, user-confirmed model analysis as weaker evidence of those choices, and other AI-written drafts or notes as records of work rather than approved judgments. A style preference in a project reference is a recorded user choice, not an externally verified fact.

## Extract and explain the changes

Read each relevant paragraph in context. Separate distinct changes rather than treating a rewritten paragraph as one issue. For each change, retain its location, initial and final wording, whether it corrects information or terminology, voice, or target-language expression, and whether it operates at word, sentence, or cross-sentence/paragraph level. Source context is needed to distinguish a translation correction from a stylistic choice.

Take the reason from the user's recorded decision or words, or from model analysis the user explicitly confirmed. Confirmed model analysis is the model's wording of the user's reason and carries less weight than the user's own words: when they conflict, follow the user's words, and weigh how strongly the recorded confirmation endorses the analysis. Mark reasons supported only by confirmed model analysis when grouping findings, and check them with the user before they enter a reference. Gather changes without a recorded reason into one concise set of questions, showing the before and after wording and labeling any possible explanation as a hypothesis. Do not infer the user's preference from the edit alone. Update the explanation after the user's response.

Group findings by the recurring mechanism across units, including recognizable source signals or target-language breaks when there are any. Leave one-off solutions in their unit records. Inspect existing reference entries for overlap, conflicts, entries that did not help, and entries previously used to resist a choice the user confirmed. Present deletion or revision candidates along with additions.

## Check candidates and propose reference updates

When a candidate would become a translation or review check, try it on a completed unit that was not used to derive it. Decide the target issues and what counts as finding them before seeing the result. Check whether the candidate locates the issue; a new wording need not match the user's final text. Keep word substitutions separate from cross-sentence and paragraph issues. Give uncertain voice judgments to the user. If an independent unit is unavailable, explain the evidence limit in the distillation record rather than adding a status label to the translator-style rule itself. A separate session can help with this check when available, but is not a prerequisite.

Route each approved durable conclusion to the document that owns it:

| Kind of conclusion | Destination |
|---|---|
| Confirmed recurring name or term | Glossary |
| Verified character, setting, or subject fact, with source | Background reference |
| A particular speaker's reading premise, voice, or translation trigger | Character or speaker profile |
| A pattern across speakers or units in interpretation or target-language writing | Translator style |

Describe the failure mode, its signal, and its scope. Avoid turning an example's exact wording into a reusable translation. Revise or remove overlapping rules instead of accumulating near-duplicates. Apply additions, revisions, and deletions only after the user approves the underlying judgment. Do not alter completed translations as part of distillation.

Record the units examined, any held-out unit used for checking, the user's decisions, reference changes, and evidence limits in the project's existing distillation record. If the project has none, create a short record only when there is an actual completed distillation to document.