byted-data-quality-review-loop

Turn the working principle "all completed task outputs must undergo strict independent review" into a Skill. Its goal is not to generate the content itself, but to force entry into 评审 -> 修改 -> 再评审 closed loop after the output is complete, until there are no issues left that need optimization.

Applicable scenarios

When preparing to deliver/submit/publish/output the final result
When the user explicitly requests "review/find faults/go through it/re-review/quality gate"
When the current executing Agent needs to perform high-standard, zero-compromise final gatekeeping on the output

Input requirements

Current output (required): can be code, documents, plans, configurations, scripts, images, audio, video, presentations, web pages, data results, file paths, links, etc.
Task context (optional but recommended): goals, constraints, audience, style requirements, existing specifications
If explicit acceptance criteria are missing, the subAgent shall complete the review criteria for this round on its own based on its own Soul and the task context, and must explicitly state this in the review report

Core process (multi-round closed loop)

The current executing Agent organizes the "current output version" and the available task context.
Through sessions_spawn start the subAgent for independent review (Round N).
The subAgent outputs a structured "Review Report" (see the fixed format below), clearly stating Pass/Fail and the issue list; if this round's review criteria were completed based on its own Soul and the task context, this must also be explicitly stated in the report.
The current executing Agent revises item by item according to the issue list, and performs necessary verification (choose an appropriate method according to the artifact type, for example: testing/comparison examples/fact checking/consistency checking/link availability and reference completeness/format and rendering checks/reproducible step verification/audio and video playback checks/subtitle and picture synchronization checks/audio clarity and noise checks/cover and asset consistency checks, etc.).
Return to step 2 and repeat review and revision until the termination conditions are met.

Termination conditions and protection limits

Default pass threshold:Blocker = 0 and Major = 0
Default maximum review rounds:3 rounds
If the limit is exceeded and it still has not passed: the current executing Agent must report the remaining issues, risks, and recommended solutions, and request the upper-level process to decide:
- Continue adding rounds, or
- Accept the risk and deliver the current version

subAgent review specifications (must be followed)

The subAgent only performs "review gatekeeping and fault finding"; it does not directly modify the output, nor write the final version on behalf of the current executing Agent.
Maintain high standards, strict requirements, and nitpicking: prioritize finding issues that would lead to failure/rollback/misleading information/increased maintenance costs.
Make no compromises: if there are items that must be fixed, the conclusion must be Fail or Pass with fixes.
Each issue must be actionable: include evidence, impact, fix suggestion, and verification method; vague generalities are prohibited.
The subAgent must be independent: judge only based on "the artifact + context + review criteria completed from its own Soul", and not rely on the current executing Agent's reasoning process.
The subAgent must switch review standards according to the artifact medium: code focuses on correctness and regression risk, documents focus on facts, structure, and readability, images focus on information accuracy and visual consistency, audio focuses on clarity, noise, sentence breaks, speaking rate, and narration errors, and video focuses on picture content, timing, transitions, subtitles, audio-visual synchronization, and completeness of information delivery.

Principle boundaries

This is a "Working Principles Skill", not a role-setting Skill. It is not responsible for defining the Agent's professional identity or personality.
Content such as roles, professional identities, and expression styles should continue to be kept in their respective SOUL.md .
This Skill is only responsible for injecting unified quality control principles: whenever there is output, it must go through independent fault-finding, multiple rounds of correction, until it meets the standard.

Severity Levels

Blocker | 阻断级: Issues that cannot be delivered or cannot pass this round of review unless fixed.
Major | 重大问题: Issues that will not immediately block, but will significantly affect quality, credibility, usability, or user understanding.
Minor | 次要问题: Issues that do not affect the main conclusion or main process, but have obvious flaws and are recommended to be fixed.
Nit | 细节优化: Non-critical optimization items such as wording, layout, rhythm, and style consistency.

Fixed Review Report Format (subAgent output template)

Review Round N

Conclusion / 结论: Pass / Pass with fixes / Fail

Severity Rule / 严重级别规则:

- Blocker = 必须修复，否则不能通过
- Major = 原则上应在通过前修复
- Minor = 建议修复
- Nit = 可选优化

Scope / 评审范围

- 本轮评审覆盖的产出类型与范围（例如：代码/文档/方案/配置/脚本/图片/音频/视频/页面；或具体文件/链接）

Must Fix / 必须修正

- [Blocker | 阻断级] 标题 / Title: ...
  证据：...
  影响：...
  修复建议：...
  验证方法：...
- [Major | 重大问题] ...

Should Fix / 建议修正

- [Minor | 次要问题] ...

Nice to Have / 可选优化

- [Nit | 细节优化] ...

Re-check Focus / 下轮复查重点

- 下一轮重点复查点（列 1-5 条）