enterprise-translation-team

name	enterprise-translation-team
description	Coordinates role-based AI agents for enterprise professional Chinese-English document translation, including terminology, first-pass translation, bilingual MQM revision, QA, context compression, and eval-driven iteration. Use when the user asks for Chinese-English business, technical, legal, product, or enterprise documentation translation workflows.
license	CC-BY-NC-SA-4.0
compatibility	Designed for GitHub Copilot CLI with custom agents and project/plugin skills.

Enterprise Translation Team

Purpose

Use this skill when an AI agent, not a human translator, must execute or design an enterprise Chinese-English document translation workflow. The goal is a controlled production line: compact context packets, role-specialized agents, structured handoffs, deterministic QA, and explicit human sign-off boundaries.

Operating rules

1. Establish evaluation criteria before iterating on prompts, skills, or agent profiles. Use evals/evals.json as the local eval manifest and references/evaluation.md for public dataset guidance.
2. Use role agents when available: translation-workflow-lead, translation-terminologist, translation-linguist, translation-reviser, and translation-qa-engineer. When loaded from this plugin, Copilot CLI may list them with the plugin namespace, for example enterprise-translation-team:translation-workflow-lead.
3. For every material step, run independent positive and negative review gates with GPT-5.5 reviewer agents. Do not proceed while either gate returns BLOCK.
4. Send each role a compact task packet instead of the whole conversation. Use references/context-packet.md as the packet template.
5. Preserve source document structure unless the user explicitly asks for adaptation. Read references/output-contracts.md before generating files.
6. Use MQM-style review findings for bilingual revision. Read references/mqm-taxonomy.md before grading errors.
7. Treat AI review as a quality gate, not a regulated human approval. Escalate legal, medical, financial, compliance, or publication-sensitive uncertainty to a human or subject-matter expert.
8. Do not write to raw\, durable wiki\, or wiki\log.jsonl merely because a translation task was performed. Only update repository knowledge when the user explicitly requests durable wiki maintenance.

Default workflow

1. Intake: identify language direction, audience, domain, risk level, files, output format, confidentiality constraints, terminology assets, and acceptance criteria.
2. Evaluation gate: choose relevant eval cases and assertions before changing skill or agent instructions.
3. Terminology pass: maintain a TBX-compatible concept-level termbase, extract job deltas, resolve conflicts, and enforce forbidden translations, acronyms, product names, placeholders, context examples, and approvals.
4. Translation or post-edit pass: translate only natural-language content while preserving Markdown, code, links, placeholders, numbers, and tables.
5. Independent bilingual revision: compare source and target, record MQM-style issues, and propose targeted fixes.
6. Post-edit and reconciliation: apply approved fixes and produce a resolved/unresolved issue report.
7. Final QA: run deterministic checks first, including termbase/schema, forbidden-term, and TBX export checks, then summarize subjective residual risks.
8. Delivery package: provide final files, termbase.job.json, termbase.delta.jsonl, termbase.tbx, terminology-review.tsv, QA report, and unresolved questions.

Minimal role handoff

When delegating to a custom agent, include only:

• Objective and owned scope.
• Source and target language direction.
• Input paths or excerpt identifiers.
• Required output paths and format.
• Glossary and style constraints relevant to that role.
• Validation command or checklist.
• Prior role outputs needed for this step.
• Explicit non-goals.

Read references/roles.md for detailed role boundaries. Read references/terminology-schema.md before creating, updating, exporting, or QA-checking terminology assets.

Available scripts

• scripts/check_translation_outputs.py validates expected eval output files, review JSON shape, termbase JSON/TBX/TSV contracts, and basic Markdown deliverables.

Run scripts from the skill directory root. Example:

mise exec -- python scripts\check_translation_outputs.py --evals evals\evals.json --case mqm-review-json --outputs path\to\outputs

Iteration rule

Do not add more instructions just because an output feels imperfect. First compare with a baseline, grade assertions with evidence, identify the smallest generalizable failure pattern, then edit the skill or agent profile.