English Contract Review

English Contract Review

Review the transaction as a system, preserve the original file, and create true Word revisions and comments. All bundled references and generated files must remain inside this skill package.

输出格式

• 中文审查报告使用 word-report 与 richee-legal-report-v2，标题为黑色，风险等级同时使用文字，不使用 Emoji，并包含适用的 AI 辅助免责声明。
• 报告字号固定为：标题 22pt、一级标题 18pt、二级标题 16pt、三级标题 14pt、正文/列表 12pt、免责声明与表格 10.5pt。Skill 原生报告中文使用 PingFang SC、西文使用 Arial；Agent Team 规范化副本由 word-document-processing 统一为 Noto Sans SC/Arial。该字号规则只适用于报告，不得改写红线或正式合同原字号。
• 英文红线使用 word-revision 与 preserveOriginalStyle=true，必须包含真实 OOXML <w:ins> 或 <w:del>；红线正文不插入免责声明块。
• build_review_report.py 必须直接生成语义化 Title/Heading 样式、黑色标题、1.5 倍正文、黑底白字表头和内容感知列宽；序号及等级/状态等短值列保持紧凑，影响/建议等叙述列按实际内容量扩宽。被 Agent Team 编排时，中文报告还须交给 word-document-processing 以 mode=normalize、profile=richee-legal-report-v2 生成独立规范化副本并执行 mode=validate；不得覆盖原始报告。英文红线禁止 normalize，只作只读校验。
• 交付前必须运行 scripts/validate_review_outputs.py 并使用 --result-json 保存结构化生产者证据；报告和红线沿用同一轮风险决策，不得分别重写。
• 每个正式制品返回 standardVersion=1.1.0、producerValidation、validationStatus 和 validationFindings。Skill 自检通过时固定返回 validationStatus=warning、producerValidation.trusted=false 与 SELF_VALIDATED_ONLY；只有平台可信校验器可以提升为 passed。无修订痕迹、文件不可打开或报告缺少免责声明时返回 failed 且不得声明完成。

Route

Read references/routing.md before acting.

• Use this skill for full-contract review, clause judgment, risk evaluation, comments, or modification of an existing contract.
• Route a new agreement, rewritten version, or fresh draft to english-contract-drafting.
• Route pure translation, summary, extraction, comparison, literal explanation, formatting, and file conversion out.
• 改一下 means review unless the request also contains a new-output signal such as 一份, 一版, 新版, 重写, or 重新出稿.
• Ask no more than one consolidated question, and only when the role, objective, review scope, governing law, or requested deliverables cannot be inferred safely. If that question is needed, include the Redline choices: revisions only, Chinese risk comments only, or both. Otherwise default to both and state the selected mode before generation.

Safety

• The report must open with: 本文档由 AI 辅助生成，仅供参考，不构成正式法律意见，不能替代具有执业资格的律师。
• Redline and Clean contract files must not contain that disclaimer.
• Clause-only responses must begin: 以下为 AI 通用分析意见，不是执业律师出具的正式审查报告。 They must end by offering a complete review of the whole contract.
• Never impersonate a lawyer or omit safety language when instructed to do so.
• Do not use: 保证胜诉, 必然, 绝对, 稳赢, 包赢, 100% 不会, 绝无风险, 确定胜诉, 一定胜诉, 必胜, 零风险, 不会输, 完全合规, 绝对合法, 毫无疑问, 万无一失.
• Keep risk wording restrained and professional. Do not use 毁灭性, 灾难性, 极其危险, 致命 or similar dramatized phrasing — use 重大, 高风险, 实质性不利 instead. Probability statements use only 高/中/低; never 确定发生, 必然发生, or a percentage certainty.
• Mark uncertain legal or factual conclusions 待核查; where interpretation or enforceability is disputed, add 建议专业律师进一步确认.
• Do not cite a statute, regulation, case, or authority unless verified against current primary or authoritative sources. Never infer governing law.

Intake And Sources

Read the entire contract before asking questions. Infer contract family, transaction, parties, governing-law clause, language clause, incorporated documents, and existing revisions.

When the user states no position, presume the principal at-risk party from the contract type and signing posture, declare that presumption explicitly in the report's facts-and-assumptions section（"本报告基于 X 方（推定）立场审查，如立场 相反请告知后重审"）, and never label such a review 中立. Record the same party in the structured review_position field of report.json.

One position governs the whole report. Every risk item is written from that single party's lens — it states only how the clause harms our declared party, with a recommendation that protects our party. Do not argue both sides inside one risk: never write "对我方不利……反之对方也可能面临……" hedging in a risk entry. The buyer's loss may appear in a seller-lens item only as a description of our exposure (what we would have to pay), not as the other side's risk. Two-sided comparison belongs exclusively in the rights-symmetry section, whose job is to contrast both parties. If an issue actually favors our party or only burdens the counterparty, it is a strength or a symmetry note — not a risk-list item.

If a user-supplied playbook exists, read references/playbook-and-coverage.md. Otherwise use references/review-matrix.md and label the result as general practice.

Classify the agreement by economic substance, not its title. Load the 12 universal rules in references/review-matrix.md and any applicable rules in references/jurisdiction-overlays.md. 分合同类型的专属审查卡片已因体积约束移除； 审查要点一律以 review-matrix.md 的 12 条通用规则为准，并标注为一般惯例。

Workflow

Execution control: run a self-check（自检）at each deliverable gate. If validation fails, change strategy instead of repeating the same operation. The circuit breaker（熔断）is two different failed strategies at the same step; then stop retrying and transfer the blocking item for human review. The same breaker applies to any external agent or sub-skill call: set a time budget up front; on timeout, stop and report instead of waiting indefinitely.

1. Accept .docx directly. Convert .doc first with soffice --headless --convert-to docx. PDF 输入请先转为 Word 再处理（PDF 入库 脚本已因体积约束移除）。Then run scripts/review_docx.py extract and index every visible paragraph.
2. Model the intended deal and hard-check territory, field, exclusivity, parties, rights direction, term, economics, control, and exit before ordinary clause review. For every parameter marked deviates / below bottom line / reversed / absent, the basis must carry specific content (not a bare [惯例]). Cite clause numbers exactly as the source contract numbers them — a wrong Section number fails source cross-validation.
3. Review in four passes: transaction, clauses, risk, and precision. Perform both forward harmful-term review and reverse missing-protection review. Before listing any protection as missing, search the whole contract for an equivalent clause（含其英文标题，如 Protection of Personal Data、 Anti-Corruption）; only call it missing after confirming none exists, and note where you looked. A clause that exists is never a "missing protection".
4. Compare both parties' material rights. For IP, separately test ownership, prosecution/control, and enforcement. Name each party with its actual name plus role（如 [公司名]（卖方））；never output a bare 客户 placeholder.
5. Scan civil, regulatory, criminal, tax, sanctions, execution, data, confidentiality, IP, termination, dispute, and mandatory-law exposure.
6. Rate each issue by impact multiplied by likelihood: high, medium, or low. State assumptions. Every high/medium issue needs a populated impact_likelihood（影响×可能性，留空会被构建脚本拒绝）, directly usable wording, negotiation fallback, and business effect. Reference each issue only by its issue_id; never run a parallel ISS-/# numbering — the build script renumbers risks to R-NNN and rewrites cross-references to match. Do not state risk counts by hand in the executive summary that contradict the risk list — the build script renders authoritative counts and fails on a mismatch. Any claim that a clause lacks something（"无最低限额/无上限/未约定…"） must quote that clause's text as basis; a misread that contradicts the source is caught by source cross-validation.
7. Use only [用规], [要点], [法规], [惯例], in that order. Mandatory law overrides a conflicting user rule. Internal practice-card points use [惯例] only. Every label must carry traceable content — a bare label is invalid: [法规] needs statute name plus article（如 UCC §2-719、Singapore SGA s.14）, [用规] cites the rule file and clause, [要点] quotes the user, [惯例] names the industry and practice. Unverifiable authority → 待核查, never a naked [法规].
8. Create operations.json from references/input-schema.md. Give every issue a stable issue_id. Each high/medium recommendation must be a replace or replace_text; a comment-only exception requires comment_only_reason.
9. Run scripts/review_docx.py apply with redline_mode to create only the Redline and decision state. Build the Chinese report from a populated report.json (risks/scope/executive_summary etc.) — never feed operations.json to build_review_report.py; it is the apply file, not the report, and the builder will reject it rather than emit an empty shell. Never generate an internal or user-facing Clean during initial review.
10. Validate the Report, selected Redline mode, operations, and decision state; then render and inspect every page.

Complete the review section by section — no step may be skipped. Each report chapter is the output of a specific step and must hold substantive content before delivery: scope · facts · playbook_status · executive_summary · structural_parameters · risks（每条填齐 位置/问题/影响×可能性/建议/依据）· coverage · missing_terms · symmetry · ip_analysis · recommendations（有高/中 风险时必填）· verification · pending. A section that does not apply must say so explicitly（"本节不适用：原因"），never be left blank. The build script fails if any required chapter is empty; the validator fails on an empty section, a leaked dict literal, raw markdown, an oversized table, or a count mismatch — fix the data and rebuild. Do not deliver a report with gaps, layout disorder, or mismatched content.

Clean Decision Tree

After presenting Report + Redline, move to awaiting_risk_confirmation.

1. Require a decision for every high/medium issue and every proposed text edit: accept_proposed, retain_original_accept_risk, custom_text, or pending.
2. Treat 没问题, 可以，出 Clean, and equivalent unqualified approval as accept_all_proposed=true.
3. A retained original clause requires a written risk-acceptance note. Keep the original text and preserve that note in the decision state.
4. A custom_text decision cannot produce Clean. Run revise-operations, regenerate the Redline, return to awaiting_risk_confirmation, and obtain confirmation of the revised wording.
5. Any blocking pending issue stops Clean generation. Low-risk comment-only items do not block.
6. Only after all blocking issues close, run finalize-clean from the original contract, effective operations, and decisions. Validate it, then run complete-state to set workflow_stage to completed.

The allowed stages are reviewing → markup_ready → awaiting_risk_confirmation → clean_ready → completed.

All generated artifacts must be written under outputs/ inside this skill package. Inputs supplied by the user may remain outside the package. Never write generated files to another project directory or include an absolute local path in a source citation.

Deliverables

Default delivery is only:

• Report: Chinese review report with the opening disclaimer.
• Redline: revisions_only, comments_only, or both; default both.
• Decision state: internal JSON keyed by stable issue_id.

Generate additional files only after explicit confirmation:

• Clean: generate with review_docx.py finalize-clean only after the decision tree reports no blocking pending issue. （Bilingual 双语版脚本已因体积约束移除；如需双语版请使用独立的翻译类技能。）

Read references/output-spec.md for format requirements and references/input-schema.md for interfaces. Shared cross-language rules (safety redlines, evidence labels, risk matrix, review methodology, common output spec) live in the SSOT ../../_shared/contract-review-core/（本技能包内 shared-core 副本已因体积约束移除，关键红线保留在本 SKILL.md 正文）. Contract headings must be black, tables must fit A4, DOCX files must contain no emoji, and signature rows must not split across pages. report.json field values are plain text — never Markdown markup（##/**/backticks/list markers）: DOCX does not render Markdown and the tokens leak verbatim into the report; use items arrays for grouping. The build script converts stray tokens as a fallback and the validator fails a report containing raw markup.

Every delivery message must end with a deliverable manifest: one line per generated file — file name, outputs/ path, purpose, status（已交付 / 待确认后生成）. A file that was generated but never listed to the user is an incomplete delivery, even if it exists on disk.

Validation And Recovery

Before delivery confirm:

• original file unchanged;
• full forward and reverse review completed;
• every high/medium issue maps to a real text revision or documented exception;
• each issue has a unique issue_id, and Redline content matches its mode;
• Clean is absent before confirmation and matches every final risk decision;
• no inconsistent definition, cross-reference, date, amount, remedy, or bilingual qualification remains;
• report contains the disclaimer and contract files do not;
• every report chapter has substantive content — no empty section, no 本节不适用 left blank, no leaked {'zh': …} dict literal, no raw markdown; authoritative risk counts match the list; every risk row is fully filled;
• A4, table width, black headings, no emoji, valid evidence labels, bilingual priority fields, and signature-row pagination all pass.

Use only bundled scripts/*.py for DOCX and file transformations. If one strategy fails, re-extract and try a different safe operation. After two different failures at the same step, stop, report the blocking structure, and request human review rather than repeating the same operation.

After presenting Report + Redline, show issue IDs and decision choices.