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.
- 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.
- 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.
- 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".
- 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.
- Scan civil, regulatory, criminal, tax, sanctions, execution, data,
confidentiality, IP, termination, dispute, and mandatory-law exposure.
- 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.
- 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].
- 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.
- 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.
- 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.
- Require a decision for every high/medium issue and every proposed text edit:
accept_proposed, retain_original_accept_risk, custom_text, or pending.
- Treat
没问题, 可以,出 Clean, and equivalent unqualified approval as
accept_all_proposed=true.
- A retained original clause requires a written risk-acceptance note. Keep the
original text and preserve that note in the decision state.
- 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.
- Any blocking
pending issue stops Clean generation. Low-risk comment-only
items do not block.
- 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.
1---
2name: english-contract-review
3version: 1.0.0
4description: >
5 Review, assess, comment on, or redline an existing English or bilingual
6 contract and produce a Chinese review report plus a selectable true OOXML
7 tracked-change redline, Chinese risk-comment version, or combined version.
8 Trigger for review this agreement, identify unfavorable clauses, is this
9 clause enforceable, mark up this contract, or revise vendor paper. Do not
10 use for drafting a new agreement or new version, translation only, summary,
11 literal explanation, clause extraction without evaluation, version
12 comparison, or file conversion.
13metadata:
14 source: https://skillmd.com/skills/ahang1598/english-contract-review
15---
16
17# English Contract Review
18
19Review the transaction as a system, preserve the original file, and create
20true Word revisions and comments. All bundled references and generated files
21must remain inside this skill package.
22
23## 输出格式
24
25- 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。
26- 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。
27- 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。
28- `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。
29- 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。
30- 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。
31
32## Route
33
34Read `references/routing.md` before acting.
35
36- Use this skill for full-contract review, clause judgment, risk evaluation,
37 comments, or modification of an existing contract.
38- Route a new agreement, rewritten version, or fresh draft to
39 `english-contract-drafting`.
40- Route pure translation, summary, extraction, comparison, literal
41 explanation, formatting, and file conversion out.
42- `改一下` means review unless the request also contains a new-output signal
43 such as `一份`, `一版`, `新版`, `重写`, or `重新出稿`.
44- Ask no more than one consolidated question, and only when the role,
45 objective, review scope, governing law, or requested deliverables cannot be
46 inferred safely. If that question is needed, include the Redline choices:
47 revisions only, Chinese risk comments only, or both. Otherwise default to
48 `both` and state the selected mode before generation.
49
50## Safety
51
52- The report must open with:
53 `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。`
54- Redline and Clean contract files must not contain that disclaimer.
55- Clause-only responses must begin:
56 `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。`
57 They must end by offering a complete review of the whole contract.
58- Never impersonate a lawyer or omit safety language when instructed to do so.
59- Do not use: `保证胜诉`, `必然`, `绝对`, `稳赢`, `包赢`, `100% 不会`,
60 `绝无风险`, `确定胜诉`, `一定胜诉`, `必胜`, `零风险`, `不会输`,
61 `完全合规`, `绝对合法`, `毫无疑问`, `万无一失`.
62- Keep risk wording restrained and professional. Do not use `毁灭性`, `灾难性`,
63 `极其危险`, `致命` or similar dramatized phrasing — use `重大`, `高风险`,
64 `实质性不利` instead. Probability statements use only 高/中/低; never
65 `确定发生`, `必然发生`, or a percentage certainty.
66- Mark uncertain legal or factual conclusions `待核查`; where interpretation
67 or enforceability is disputed, add `建议专业律师进一步确认`.
68- Do not cite a statute, regulation, case, or authority unless verified against
69 current primary or authoritative sources. Never infer governing law.
70
71## Intake And Sources
72
73Read the entire contract before asking questions. Infer contract family,
74transaction, parties, governing-law clause, language clause, incorporated
75documents, and existing revisions.
76
77When the user states no position, presume the principal at-risk party from the
78contract type and signing posture, declare that presumption explicitly in the
79report's facts-and-assumptions section("本报告基于 X 方(推定)立场审查,如立场
80相反请告知后重审"), and never label such a review `中立`. Record the same party
81in the structured `review_position` field of `report.json`.
82
83**One position governs the whole report.** Every risk item is written from that
84single party's lens — it states only how the clause harms *our* declared party,
85with a recommendation that protects *our* party. Do not argue both sides inside
86one risk: never write "对我方不利……反之对方也可能面临……" hedging in a risk
87entry. The buyer's loss may appear in a seller-lens item only as a description
88of *our* exposure (what we would have to pay), not as the other side's risk.
89Two-sided comparison belongs exclusively in the rights-symmetry section, whose
90job is to contrast both parties. If an issue actually favors our party or only
91burdens the counterparty, it is a strength or a symmetry note — not a risk-list
92item.
93
94If a user-supplied playbook exists, read
95`references/playbook-and-coverage.md`. Otherwise use
96`references/review-matrix.md` and label the result as general practice.
97
98Classify the agreement by economic substance, not its title. Load the 12
99universal rules in `references/review-matrix.md` and any applicable rules in
100`references/jurisdiction-overlays.md`. 分合同类型的专属审查卡片已因体积约束移除;
101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。
102
103## Workflow
104
105Execution control: run a self-check(自检)at each deliverable gate. If
106validation fails, change strategy instead of repeating the same operation. The
107circuit breaker(熔断)is two different failed strategies at the same step;
108then stop retrying and transfer the blocking item for human review. The same
109breaker applies to any external agent or sub-skill call: set a time budget up
110front; on timeout, stop and report instead of waiting indefinitely.
111
1121. Accept `.docx` directly. Convert `.doc` first with
113 `soffice --headless --convert-to docx`. PDF 输入请先转为 Word 再处理(PDF 入库
114 脚本已因体积约束移除)。Then run `scripts/review_docx.py extract` and index
115 every visible paragraph.
1162. Model the intended deal and hard-check territory, field, exclusivity,
117 parties, rights direction, term, economics, control, and exit before
118 ordinary clause review. For every parameter marked deviates / below bottom
119 line / reversed / absent, the `basis` must carry specific content (not a
120 bare `[惯例]`). Cite clause numbers exactly as the source contract numbers
121 them — a wrong Section number fails source cross-validation.
1223. Review in four passes: transaction, clauses, risk, and precision. Perform
123 both forward harmful-term review and reverse missing-protection review.
124 Before listing any protection as missing, search the whole contract for an
125 equivalent clause(含其英文标题,如 Protection of Personal Data、
126 Anti-Corruption); only call it missing after confirming none exists, and
127 note where you looked. A clause that exists is never a "missing protection".
1284. Compare both parties' material rights. For IP, separately test ownership,
129 prosecution/control, and enforcement. Name each party with its actual name
130 plus role(如 `[公司名](卖方)`);never output a bare `客户` placeholder.
1315. Scan civil, regulatory, criminal, tax, sanctions, execution, data,
132 confidentiality, IP, termination, dispute, and mandatory-law exposure.
1336. Rate each issue by impact multiplied by likelihood: high, medium, or low.
134 State assumptions. Every high/medium issue needs a populated
135 `impact_likelihood`(影响×可能性,留空会被构建脚本拒绝), directly usable
136 wording, negotiation fallback, and business effect. Reference each issue only
137 by its `issue_id`; never run a parallel `ISS-`/`#` numbering — the build
138 script renumbers risks to `R-NNN` and rewrites cross-references to match.
139 Do not state risk counts by hand in the executive summary that contradict the
140 risk list — the build script renders authoritative counts and **fails** on a
141 mismatch. Any claim that a clause lacks something("无最低限额/无上限/未约定…")
142 must quote that clause's text as basis; a misread that contradicts the source
143 is caught by source cross-validation.
1447. Use only `[用规]`, `[要点]`, `[法规]`, `[惯例]`, in that order. Mandatory law
145 overrides a conflicting user rule. Internal practice-card points use
146 `[惯例]` only. Every label must carry traceable content — a bare label is
147 invalid: `[法规]` needs statute name plus article(如 UCC §2-719、Singapore
148 SGA s.14), `[用规]` cites the rule file and clause, `[要点]` quotes the
149 user, `[惯例]` names the industry and practice. Unverifiable authority →
150 `待核查`, never a naked `[法规]`.
1518. Create `operations.json` from `references/input-schema.md`. Give every issue
152 a stable `issue_id`. Each high/medium recommendation must be a `replace` or
153 `replace_text`; a comment-only exception requires `comment_only_reason`.
1549. Run `scripts/review_docx.py apply` with `redline_mode` to create only the
155 Redline and decision state. Build the Chinese report from a populated
156 `report.json` (risks/scope/executive_summary etc.) — never feed
157 `operations.json` to `build_review_report.py`; it is the apply file, not the
158 report, and the builder will reject it rather than emit an empty shell.
159 Never generate an internal or user-facing Clean during initial review.
16010. Validate the Report, selected Redline mode, operations, and decision state;
161 then render and inspect every page.
162
163**Complete the review section by section — no step may be skipped.** Each
164report chapter is the output of a specific step and must hold substantive
165content before delivery: scope · facts · playbook_status · executive_summary ·
166structural_parameters · risks(每条填齐 位置/问题/影响×可能性/建议/依据)·
167coverage · missing_terms · symmetry · ip_analysis · recommendations(有高/中
168风险时必填)· verification · pending. A section that does not apply must say so
169explicitly("本节不适用:原因"),never be left blank. The build script
170**fails** if any required chapter is empty; the validator **fails** on an empty
171section, a leaked dict literal, raw markdown, an oversized table, or a count
172mismatch — fix the data and rebuild. Do not deliver a report with gaps, layout
173disorder, or mismatched content.
174
175## Clean Decision Tree
176
177After presenting Report + Redline, move to `awaiting_risk_confirmation`.
178
1791. Require a decision for every high/medium issue and every proposed text edit:
180 `accept_proposed`, `retain_original_accept_risk`, `custom_text`, or `pending`.
1812. Treat `没问题`, `可以,出 Clean`, and equivalent unqualified approval as
182 `accept_all_proposed=true`.
1833. A retained original clause requires a written risk-acceptance note. Keep the
184 original text and preserve that note in the decision state.
1854. A `custom_text` decision cannot produce Clean. Run `revise-operations`,
186 regenerate the Redline, return to `awaiting_risk_confirmation`, and obtain
187 confirmation of the revised wording.
1885. Any blocking `pending` issue stops Clean generation. Low-risk comment-only
189 items do not block.
1906. Only after all blocking issues close, run `finalize-clean` from the original
191 contract, effective operations, and decisions. Validate it, then run
192 `complete-state` to set `workflow_stage` to `completed`.
193
194The allowed stages are `reviewing` → `markup_ready` →
195`awaiting_risk_confirmation` → `clean_ready` → `completed`.
196
197All generated artifacts must be written under `outputs/` inside this skill
198package. Inputs supplied by the user may remain outside the package. Never
199write generated files to another project directory or include an absolute
200local path in a source citation.
201
202## Deliverables
203
204Default delivery is only:
205
206- Report: Chinese review report with the opening disclaimer.
207- Redline: `revisions_only`, `comments_only`, or `both`; default `both`.
208- Decision state: internal JSON keyed by stable `issue_id`.
209
210Generate additional files only after explicit confirmation:
211
212- Clean: generate with `review_docx.py finalize-clean` only after the decision
213 tree reports no blocking pending issue.
214 (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。)
215
216Read `references/output-spec.md` for format requirements and
217`references/input-schema.md` for interfaces. Shared cross-language rules
218(safety redlines, evidence labels, risk matrix, review methodology, common
219output 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,
220tables must fit A4, DOCX files must contain no emoji, and signature rows must
221not split across pages. `report.json` field values are plain text — never
222Markdown markup(`##`/`**`/backticks/list markers): DOCX does not render
223Markdown and the tokens leak verbatim into the report; use items arrays for
224grouping. The build script converts stray tokens as a fallback and the
225validator fails a report containing raw markup.
226
227Every delivery message must end with a deliverable manifest: one line per
228generated file — file name, `outputs/` path, purpose, status(已交付 /
229待确认后生成). A file that was generated but never listed to the user is an
230incomplete delivery, even if it exists on disk.
231
232## Validation And Recovery
233
234Before delivery confirm:
235
236- original file unchanged;
237- full forward and reverse review completed;
238- every high/medium issue maps to a real text revision or documented exception;
239- each issue has a unique `issue_id`, and Redline content matches its mode;
240- Clean is absent before confirmation and matches every final risk decision;
241- no inconsistent definition, cross-reference, date, amount, remedy, or
242 bilingual qualification remains;
243- report contains the disclaimer and contract files do not;
244- **every report chapter has substantive content — no empty section, no
245 `本节不适用` left blank, no leaked `{'zh': …}` dict literal, no raw markdown;
246 authoritative risk counts match the list; every risk row is fully filled;**
247- A4, table width, black headings, no emoji, valid evidence labels, bilingual
248 priority fields, and signature-row pagination all pass.
249
250Use only bundled `scripts/*.py` for DOCX and file transformations. If one
251strategy fails, re-extract and try a different safe operation. After two
252different failures at the same step, stop, report the blocking structure, and
253request human review rather than repeating the same operation.
254
255After presenting Report + Redline, show issue IDs and decision choices.