content-reviewer
Skill1 starsChanged 27 days agoApache-2.0View on GitHub
Review generated commerce videos for product mismatch, human or subject artifacts, motion discontinuity, and human/product proportion anomalies. Use after generated segments are downloaded, before editing, and again after final rendering when Codex must identify evidence-backed time ranges and decide whether to keep, trim, regenerate, block, or mark the final video publish-ready.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- TeqfocusAdmin/ClaudeSkillsRepotheir own version, first seen Sep 22, 2026
- genfeedai/genfeed.aitheir own version, first seen Sep 12, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- genfeedai/genfeed.ai★ 4same name
- TeqfocusAdmin/ClaudeSkillsRepo★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository10 other files it gives its agents.
- creative-runnerSkill
- douyin-music-librarySkill
- ecommerce-operationsSkill
- loopilot-deployerSkill
- loopilot-doctorSkill
- loopilot-workbenchSkill
- reusable-material-librarySkill
- video-content-understandingSkill
- video-editor-runnerSkill
- video-script-generatorSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 5 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 16
Use this skill as the objective quality evidence and routing layer between generation, editing, and final delivery.
- Distinctive: line 59
Treat perspective, motion blur, occlusion, and wide-angle distortion as uncertainty unless adjacent frames prove a real anomaly.
- Very specific: line 61
Use original product_source.json images as the product fact source.
- Distinctive: line 78
Do not overwrite source media.
- Distinctive: line 80
Do not call video generation, editing, publishing, or external write workflows.
- Very specific: line 81
Do not convert Seed candidates directly into remove_ranges.
- Distinctive: line 82
Do not use image generation or generative image editing for formal issue evidence.
- Very specific: line 83
Do not produce passed or publish_ready unless every required image was loaded and reviewed.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
创意返工按相邻 创意作者交接 返回具体问题、对应镜头/时间段和证据。主会话负责验收并判断创意缺口、媒体未实现或剪辑/素材问题,审核工具辅助取证,不默认另派验收子 Agent。创意修改交原作者,作者只交修订成稿;主会话按任务分别判断故事推进、宠物主动性、有效互动与用途表达等适用要求,脚本验收或技术通过不替代实际成片验收。
任务方案与交付
当前运行 Agent 在本阶段前读取相邻 ecommerce-operations/references/task-execution.md,沿用本任务最新有效方案和适用偏好,按真实结果交付并记录成功操作。已在同一任务读取时直接复用;不反向重启总控或重复业务动作。Agent 主导顺序、分支与恢复,本 Skill 返回实际结果和正式恢复线索。该参考暂不可读时说明缺口并继续当前可执行工作,不新增文档或格式硬门槛。
Content Reviewer
Use this skill as the objective quality evidence and routing layer between generation, editing, and final delivery. It reviews and recommends; it never edits, regenerates, publishes, or independently vetoes a plan-driven execution.
product_asset_remix_v2 使用方案驱动审核入口。先运行 prepare_content_review.py --plan-driven-remix-manifest ... --video ... --output-dir ...,由 reviewed 方案决定是否需要 AI 审核;无论是否启用 AI,都必须有绑定最终视频 SHA 的 plan-driven-remix-local-validation-v1。再运行 finalize_content_review.py --plan-driven-remix --request ... --local-validation ... [--model-result ...] --output ...。商品一致性、方案遵从、类目宣称合规和媒体质量全部通过时才写 formal_publish_ready=true;该值只表示本地成片审核通过,不授予发布权限。
对 plan_driven_autonomous_v1,finalize_quality_gate_v2() 保留原始通过/失败、证据与 execution_advice。执行模型通过所属 Skill,在已有授权与方案预算内选择本地修复或重新生成;具备权限的人可对精确绑定的最终成片直接完成 execution。人工覆盖不得改写自动质检事实,也不得授权发布或其它外部写入。
视频复刻必须加载 hash 绑定的 product_identity_lock.json 并对每个 feature_id 输出 product_feature_checks[]。品牌文字、核心图案和位置关系不能用“非核心细节”驳回;持续或跨镜头重复错配只能 regenerate_required,证据不足只能 blocked。最终成片还必须回读 source_review_lineage[],上游负向商品结论不能通过裁剪、放大或遮挡清除。
新视频复刻 revision 使用两个业务门禁:生成素材完成后执行 material_entry,最终渲染及每次修改后执行完整 final_delivery。两者都消费 LP-063 的同一 video-evidence-ref-v1,重新计算业务结论,不复制缓存。必须同时检查证据完整、物理可信、商品口播一致、真实镜头时长、源区间唯一、语义重复、素材充足、时间轴连续和全片覆盖;任一 uncertain 先补指定区间证据,仍不确定则 blocked。
生成片段的 material_entry 默认调用 video-content-understanding prepare-local-review --account-key,以零 Provider 方式取得当前片段的正式 evidence ref;不得为此触发参考视频理解 Provider。各片段独立完成 V2 视觉审核后,才可用 scripts/aggregate_segment_reviews.py --segment-report ... 汇总连续 approved_timeline。汇总器只合并已通过的真实分段报告,校验每段媒体 path/SHA/evidence ref,不生成或改判质量结论。
镜头长度、beat 覆盖、同源复用、置信度、证据张数和删减比例交由模型结合任务、音画与实际证据判断;不得用固定阈值改写模型结论。真实时间码、媒体/证据来源和当前任务要求继续校验。最终 SHA 变化后重审完整成片;普通 publish_ready 不替代 TTS 音频完整性或适用的 reference fidelity,也不授权发布。
同一 final media、edit manifest、商品事实、方案 revision、review profile 与 sampling 只允许一份 active 成片审核。准备入口把完整 request 计算为 review_fingerprint;已存在且由 content-review-completion-receipt-v1 精确绑定的 request/model result/report 直接返回 already_completed,不再次请求模型。任一输入变化不得覆盖旧报告,必须进入新指纹工作区;旧显式 output dir 保持只读兼容。
新局部 TTS 正式资产接受 continuous-tts-manifest-v4,并回读源 manifest、revision receipt、piecewise mapping、透明字幕布局、Provider 字幕和 tts-audio-validation-v1。历史 V3 仅只读兼容。定向画面返修必须回读 video-repair-evidence-v1;PCM 相等只继承音频完整性事实,不能继承画面审核,最终普通审核和参考保真始终针对完整新媒体重跑。
商品原创审核
creation_mode=product_original 复用 segment/final 本地 FFmpeg 取证和 Codex 逐镜审核。兼容参数 --replication-dir 指向原创任务目录,读取原生 original_storyboard.json、原商品来源和补料,不补造参考结构、人物模板或 reference-fidelity。
商品/主体一致性、场景动作、音画文字、媒体技术质量与原创方案符合性分别记录;不要将生成图片当商品真值。审核负向事实保留,由执行模型在批准范围内恢复;内容通过不授权费用或发布。最终交付仍需 exact media/plan 绑定的 finalizer,报告直接回到原创作中心任务。
女装普通单视频流程
女装普通单视频审核时,读取 womenswear-workflow.md。同一任务兼有多个分支时读取对应分支;其中的原有条件、顺序和例外继续适用。
Segment workflow
分段视频审核时,读取 segment-workflow.md。同一任务兼有多个分支时读取对应分支;其中的原有条件、顺序和例外继续适用。
Final-render workflow
最终成片审核时,读取 final-render-workflow.md。同一任务兼有多个分支时读取对应分支;其中的原有条件、顺序和例外继续适用。
Review rules
- 视频复刻 Segment / final 只使用 FFmpeg 本地候选证据和 Codex;Seed 2.1 Pro 只保留给参考视频完整理解。女装 direct review 与历史兼容入口继续使用既有 Seed 候选合同。
review_sampling只允许overview_fps / action_fps / supplement_fps / max_codex_images,解析顺序为 content-reviewer 默认值、creation mode 匹配的 creative profile、任务 override。- Codex 输入保留 timeline contact sheet、每镜代表帧和必要商品细节,默认每批 24 张;证据超过一批时继续分批读取,不丢弃镜头或设置累计上限。尺寸为生成时的优化默认,原 consumer 不因图幅或拼图格数拒绝有效证据。
- Seed output is not the final verdict. Codex must inspect the extracted frames before confirming an issue.
- Review every shot even when Seed returns no candidates.
- Treat perspective, motion blur, occlusion, and wide-angle distortion as uncertainty unless adjacent frames prove a real anomaly.
- Allow intentional hard cuts and jump cuts in womenswear product-display videos. When the edit manifest confirms a cut and adjacent evidence shows consistent product features, normal human structure, and no penetration or impossible interaction, do not confirm
action_jumpsolely because pose, bag-carrying hand, viewpoint, or action position changes across the cut. - Use original
product_source.jsonimages as the product fact source. Generated references and video frames cannot replace them. - Require real evidence paths and valid timecodes for every confirmed issue.
- Require deterministic local annotation evidence for every confirmed issue. The renderer may draw only on copies and must preserve source hashes.
- 缺真实证据或结论时返回具体缺口,由执行模型补读、修订并重入原审核入口。置信度数值不单独决定停机;模型明确 uncertain/failed 必须保留,不能补造 passed。
Decisions
assembly_required: clean segment that still needs assembly.passed: clean standalone segment.edit_required: every issue is short, precisely removable, and removal preserves required product display, action, and speech semantics.regenerate_required: 实际模型确认的持续商品错误或不可局部修复的语义/画面问题;不按固定删减比例或剩余秒数替模型裁决。publish_ready: clean final render after post-edit review.reference_fidelity_regeneration_required: 普通质量通过,但当前复刻模式定义的原始画面硬门禁未达标;诊断项不得单独触发该决定。blocked: 当前缺证据或模型无法确认;由原执行模型补证据/恢复,非全任务永久终态。
Boundaries
- Do not overwrite source media.
- Seed 初审需要 Provider 时必须显式提供
--ark-provider-profile的 v3 合同;生产 Key 仅从 Keychain 解析并在隔离子进程注入。命令禁止--profile/--api-key,缺 Key 或绑定漂移时零调用,401/403 不重试且不得回退 SSO。 - Do not call video generation, editing, publishing, or external write workflows.
- Do not convert Seed candidates directly into
remove_ranges. - Do not use image generation or generative image editing for formal issue evidence.
- Do not produce
passedorpublish_readyunless every required image was loaded and reviewed.
The preparation step writes an active reports/report_manifest.json; finalize atomically replaces it with the terminal decision and account/run/product identity. Final publish_ready is review evidence only. It can enter the publisher's local queue only through the separate evidence-checking handoff and is never platform authorization.
同商品混剪终审
creation_mode=product_asset_remix_v1 只允许 final-stage V2 审核。准备阶段使用本地 FFmpeg 证据生成 local-final-evidence-v1,必须绑定最终视频 SHA256、原始 local-review-evidence-v1、逐镜 artifact,且 provider=local_ffmpeg / seed_call_count=0;不得伪装成或覆盖源素材的 video-evidence-ref-v1。
终审继续执行九项质量门,并额外回读 V5 多源 approved_timeline 与 source_review_lineage[]:每个 atom 的源 SHA、整数 PTS、允许区间和唯一 origin mapping 必须闭包。任一源审核失败、区间越界、hash 漂移、证据不完整或质量门 uncertain 时阻断;全部通过才返回 publish_ready,但该结论仍不构成发布授权。
条件式屏显文字最终审核
最终 prepare 从实际 edit manifest 生成 screen-text-review-contract-v1,逐层检查文字依据、时序、可读性、遮挡与方案符合性;omit 须确认没有意外新增文字。发布短标题独立存在,不要求补造卖点轨。文字失败交回既有 repair_screen_text_track,允许有新输入的多轮零 Provider 修订;修订后重渲染/重审当前媒体,不重新生成已可用视频。finalizer 回读当前 decision、resolution、render/preserve、review 与适用的 repair receipt,不抹除原失败事实。
Resources
- Input/output contract:
references/input-output-contract.md - Runtime profile:
templates/content-review-profile.json - Preparation and Seed review:
scripts/prepare_content_review.py - Deterministic validation and routing:
scripts/finalize_content_review.py - Semantic preservation evaluator and plan builder:
scripts/semantic_preservation.py - Post-trim semantic finalizer:
scripts/finalize_semantic_preservation.py - Deterministic issue annotation renderer:
scripts/render_issue_evidence.py
原生事实交接
审核结果通过原结构校验、九项质量检查及决定重算后直接输出,不登记中央兼容回执。模型按需恢复输入表达差异,但不得改变 issues、商品一致性、reference fidelity、技术质量或最终审核事实。失败与 uncertain 交回原审核/创作能力;转换成功不等于审核通过。
混剪的真实 failed/uncertain 决定继续保留,Agent 根据具体时间窗返剪或补证后生成新媒体/新审核输入,不把格式适配写成通过。V2 finalizer 同步原工作台报告,formal_publish_ready 来自本次审核;附加交付待办由任务 Agent 另行回读,不因审核通过自动视为所有交付物齐备。
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

