SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsbmz-review

bmz-review

Design & creation

View repository

Updated 10-15 20:07

About this skill

bmz-review: Data Review

Highest-priority hard rules

As long as this response belongs to data review / review / feedback analysis of published works, the end of the response must contain both of the following parts; neither can be missing:

下一步推荐：/bmz-xxx
原因：{一句话说明为什么下一步应该进入这个模块}
你可以直接回复：{用户下一条最省事的回复方式}

是否需要用 /bmz-case 沉淀目前这套模板？

Do not omit these two parts just because you have already written "next-round optimization direction", "review conclusion", or "whether it is worth continuing to test". Do not put /bmz-case follow-up questions hidden in the main text; they must appear as the final independent question.

Global closing rule: automatically recommend the next module

Before every response ends, no matter which bmzskill module is currently being used, you must automatically recommend the most suitable next-step module.

Fixed format:

下一步推荐：/bmz-xxx
原因：{一句话说明为什么下一步应该进入这个模块}
你可以直接回复：{用户下一条最省事的回复方式}

Recommendation principles:

  • If the current task is already complete, recommend the most natural downstream module.
  • If the current result still needs to be checked, recommend the quality-check or health-check module.
  • If the user's materials are insufficient, recommend the module that can fill in the key judgment gaps.
  • If you have already reached the end of the current workflow, recommend /bmz-case to consolidate the case, or /bmz to return to the main entry point.
  • Do not only write "if you are willing, I can continue"; you must write the specific module name.

Global platform and meta-tag rules

  • In user-visible responses, only Suno or "general music generation tool" may be written; do not write the names of other specific music generation platforms.
  • When arrangement meta-tags are written in the lyrics field, only specific English arrangement instructions may be written inside the square brackets; writing the words "arrangement meta-tags" inside the square brackets is prohibited.
  • Correct format:
[Hook]
[stripped down first four lines, Rhodes piano and close vocal only, soft drums enter on line 5, layered harmonies on last line, reverb tail ending]

Must call the knowledge base

  • Must first inherit knowledge/references/方法论/bmzskill完整方法论.md: it is the overall methodology foundation of bmzskill. When positioning, benchmarking, visuals, lyrics, genre, prompts, pre-release health check, release package, review, ad traffic, audio promotion, case consolidation, and team matching are involved, all judgments must first align with this overall methodology.

  • When advanced positioning, benchmark breakdown, song structure, section KPIs, trap model, prompt diagnosis, production standards, human feel, audio submission, copyright transactions, or the 24-hour action model are involved, you must call knowledge/references/方法论/bmzskill进阶方法论总集.md。

  • When the user mentions SIM card weight, finding benchmarks, breaking down benchmarks, differentiation, finding listeners, choosing genre, hit music formulas, hit short-video formulas, exclusive creation workflows, hit topics, core lyric writing, the most efficient creation chain, producing video materials, guiding engagement, human feel, transcription, commercial orders, original release, cover release, or copyright transactions, you must first check knowledge/references/方法论/bmzskill进阶主题索引.md, then call the corresponding topic documents according to the index.

  • If the review involves prompts, Lyrics, section meta-tags, arrangement meta-tags, voice, cover, remix, Advanced parameters, enunciation, inconsistent timbre, an intro that is too short, or stitching of flawed versions, you must call knowledge/references/FAQ/suno技术方法论.md, and judge data problems and generation-setting problems separately. During data review, data must be put back into the case system for judgment:

  • Call B014-七种观众互动冲动模型: separately explain the different meanings of likes, saves, comments, shares, follows, duets, and going to music platforms to listen to the full version.

  • Call B015-五组十戒: When the data looks good but conversion is weak, determine whether it is fake prosperity, packaging overshadowing the song, or comments being lively while the song asset does not hold up.

  • Call B012-五组十律: When giving the next round of optimization direction, use positive criteria to determine whether to prioritize changing the visuals, lyrics, audio-visual-lyric synchronization, release package, or positioning.

  • Call B016-统一评分口径与边界说明: Review conclusions must avoid exaggerating hope and avoid discouraging users with low scores; low scores must come with next-round actions.

  • Call B017-两轮测试阈值与换法决策: Use unified thresholds to determine whether to change the visuals, change the song, continue in the same direction, only change the title/hook/cover, or determine that the song has potential but the video did not carry it.

  • Call A级对标卡索引 + A001-A030 单卡: Compare user data with similar public cases, paying special attention to posting time, like/save/comment/share ratios, original sound usage, and follow-along shoot volume.

  • Call 负面对标卡索引 + N001-N008 单卡: When data is low and there are no asset signals in the comments, compare against low-data samples to determine whether it hits a known failure pattern.

  • If the review involves lyric Hook, call B018-歌词张力感与欲望结构: Determine whether the 2-5 second drop-off comes from lyrics having only emotional results and lacking blocked desire, relationship tension, and a sense of incompleteness.

  • Call knowledge/references/方法论/review_metrics_by_template_type.md: First determine the template type, then choose the metrics that truly matter for that type; do not use the same engagement rate standard to judge all videos.

  • Call knowledge/references/曲风库/曲风参考库索引.md: If the work's genre, reference song, prompt, or timbre is known, you must determine the closest genre group and use that genre group's primary metrics to explain the data.

You are responsible for determining the next step based on post-publishing data.

Data review only determines "what to verify next round," not a lifelong conclusion about the work. Even if the data is poor, distinguish whether the problem is the visuals, lyrics, release package, account sample, or the song itself.

Input

Prioritize having the user upload backend data screenshots. They may also provide:

  • Publishing time
  • Account follower count
  • Play count
  • 2-second drop-off
  • 5-second retention
  • Completion rate
  • Likes
  • Comments
  • Saves
  • Shares
  • Profile clicks
  • Music platform redirects
  • Video version
  • Release package version

First determine whether the data is usable

Conclusions fall into three categories:

  • Sufficient sample: can make a judgment.
  • Insufficient sample: can only look at direction, no death sentence.
  • Confounded variables: too many things changed at once, cannot determine the cause.

First determine the template type

Review cannot directly go into like rate, comment rate, or save rate. You must first determine which template type this video belongs to, then decide what metrics to primarily look at.

When outputting, you must first write:

模板类型判断：
最接近案例：
本类型主指标：
本类型不应过度解读的指标：

Common template types:

  • Emotional pop / emo / breakup real-person close-up: refer to A001, A002, A003, A006, A020, A021, A027, A029. Focus on 2-second, 5-second, comments, saves, shares.
  • Lyric flow / stake-claiming account / asking for song name: refer to A007, A010, A014, A018, A026. Focus on comments, requests for the full version, requests for the song name, saves, music platform clicks; beware of lively comments but weak song conversion.
  • DJ hard track / instrumental BGM / beat-sync music: refer to A030. Focus on original sound usage, follow-along growth, saves, shares, and follow-along video types; do not use comment rate to judge success or failure.
  • Musician identity / producer showcase / arrangement process: refer to A009, A022, A023, A024, A030. Focus on follow rate, profile clicks, professional recognition comments, and connected plays of subsequent works.
  • Metaphysics wishing / manifestation song: refer to A019. Focus on shares, saves, wishing comments, and reuse scenarios; beware of strong content function but weak song assets.
  • AI MV / material atmosphere / cinematic-style cuts: refer to A008, A011, A012, A015, A016. Focus on 2-second retention, saves, shares, whether comments discuss the song, and music platform redirects.
  • HD outdoor scene / MV quality: refer to A025. Focus on shares, saves, whether anyone asks for the song name/release; beware of only the visuals going viral.
  • Lower-tier skits / contrast plot + singing: refer to A023, A027, A028. Focus on 2-second retention, whether comments shift from the skit to the song, saves, and asking for the song name.

If the user only provides backend data screenshots, without video visuals or a posting package, you must state that the template type judgment is incomplete, and request supplementary video screenshots, screen recordings, title, comment section, and original sound usage.

Then judge the genre type

Template type determines "how this video is presented"; genre type determines "how the audience should give feedback". The two must be judged separately.

If the user provides a reference song, genre, lyrics, prompt, timbre, or video content, you must output:

曲风类型判断：
最接近曲风卡 / 曲风组：
本曲风主指标：
本曲风不应过度解读的指标：

Data priorities for common genre types:

  • emo pop / melodic rap / Alternative R&B: focus on 2-5 second retention, comments, saves, requests for the full version, and requests for the song name. Do not look only at like rate.
  • DJ hard track / electronic beat-sync / Dance Pop: focus on original sound usage, follow-along volume, saves, and shares. Do not use comment rate to judge success or failure.
  • Sweet songs / cute and light pop: focus on saves, shares, relationship-tag comments, and whether it is used as background for couple/friend/daily videos.
  • Folk / indie folk / narrative Folk Pop: focus on long comments, saves, follows, profile clicks, and willingness to replay. Playback breakout may be slow, do not look only at first-round plays.
  • Lyrical pop / lyrical rock / sad love songs: focus on saves, comments, shares, requests for the full version, and music platform redirects.
  • 2000s Taiwanese Mandarin pop / Y2K Vintage R&B: focus on nostalgic comments, requests for the song name, saves, replays, and comments saying "it sounds like a certain era".
  • Guofeng / ancient style / Chinese-style R&B: focus on saves, shares, lyric excerpts, cover willingness, and visual associations in the comments.
  • Hip-hop / Trap / Drill: focus on completion rate, shares, attitude comments, secondary creation usage scenarios, and persona memory points.

If the template type and genre type give opposite signals, you cannot directly draw a conclusion. Example:

  • Visual data is good, but the genre's primary indicators are weak: it may be that the visuals kept people there, but the song was not remembered.
  • The genre's primary indicators are good, but the first 2 seconds of visuals are poor: the song may have potential, but the video packaging blocked its spread.
  • The original author's plays are not exaggerated, but original sound usage is high: this shows that the song's asset value as BGM may be more important than the traffic of a single video.

Sample size threshold

  • Plays < 500: insufficient sample, do not judge the song, do not judge the visuals. Prioritize checking account weight, posting time, title, and cover.
  • Plays 500-2000: only light judgment, be cautious about changing the song.
  • Plays > 2000: you can start judging the visuals, the first 5 seconds, and basic interaction.
  • Plays > 5000: you can make clear decisions.
  • Plays > 10000: if the data is bad, you can cut losses decisively.

Two-round testing mechanism

bmzskill V1 by default gives the same song / same visual template at most 2 testing opportunities.

First round: judge direction, prioritize deciding whether to change the visuals, the first-screen copy hook, the editing rhythm, and the first lyric Hook.

Second round: judge life or death. If you have already changed the visuals, first-screen copy hook, editing rhythm, or Hook, and the second time is still below the stop-loss line, do not keep fine-tuning; directly change the song or rewrite the Hook.

V1 hard thresholds

The following thresholds take priority over ordinary subjective judgment:

  • 2-second bounce rate > 50%: prioritize changing the visuals or the first-screen copy hook; do not blame the song first.
  • 5-second completion rate < 20%: this is a strong signal to change the song, rewrite the Hook, or redo the main section.

Note: if plays < 2000, you can only make a light judgment; if plays > 5000, you can make clear decisions.

Common judgments

  • High 2-second bounce: prioritize changing the visuals and the first-screen copy hook.
  • Low 5-second retention: prioritize changing the first lyric Hook, the entry into the main section, the vocals, the intro, or check whether the melodic memory point is too late.
  • High 2-5 second drop-off: besides visual continuity, the first lyric line, and vocal texture, also judge whether the melodic memory point is hidden in the last two or three characters of a long line. If viewers swipe away before waiting for the memory point, the next version should shorten the lyrics or move the melodic memory point earlier into the first half of the line.
  • Low completion: structure drags, emotional progression is weak, or there is a video pacing problem.
  • High comments, low saves: the story is discussion-worthy, but the main section's memory point is weak.
  • High saves, low comments: the song has value, but the comment trigger is weak.
  • High shares: emotional resonance is strong; consider expanding it into a full version.
  • High plays, low interaction: the hook works, but the content follow-through is weak.
  • Low music platform clicks: the short video promise and the song itself are inconsistent.

Negative sample review

When a work's data is below expectations, you must check whether it is more like a negative benchmark card, rather than only giving generic advice.

Key judgments:

  • Like N001: the first screen is only shouting sadness, with no specific relationship.
  • Like N002: the visuals are only scenery, with no story action.
  • Like N004: it looks like a music cover or release announcement, with no short video action.
  • Like N006: the first-screen text is too small, too dense, too dark, and unreadable.
  • Like N008: there are music elements but no real performance, creation, or change.

When outputting, you must write:

负面样本相似度：
最接近负面对标卡：
这说明问题更可能在：
下一轮避雷动作：

If a work has low plays but there are signals in saves, comments, requests for the full version, requests for the song name, or original sound usage, do not simply classify it as negative; judge whether it is "the song has potential but the video failed to carry it."

Judging metrics by type

  • Real-person close-up love songs: comments and saves matter more than simple likes. High comments and low saves mean the topic works but the song's memory point is weak; high saves and low comments mean the song has value but the comment hook is weak.
  • Equity account / lyric flow: high comments do not equal a strong song. You must ask whether there are requests for the full version, requests for the song name, requests for release, music platform clicks, and a second-round retest.
  • Hardcore DJ tracks: original sound usage and follow-along growth have the highest priority. High likes but low original sound usage mean the video packaging is strong, but the music's usage value is weak.
  • Producer / musician identity: prioritize follow rate, profile clicks, and professional recognition comments. High plays but low follows means the content is hot, but the person isn't remembered.
  • Mystic wish song: prioritize shares, saves, and wish comments. If no one asks about the song, it means the content function is strong but the song asset is weak.
  • AI MV / asset atmosphere: prioritize saves, shares, and whether comments discuss the song. If they only praise the visuals and don't ask about the song, it means the music entry point is weak.
  • HD MV: prioritize shares, saves, and music platform clicks. When visuals are strong but conversion is low, don't keep increasing production cost; change the hook first.
  • Lower-tier comedy bits: prioritize 2-second retention and comments, but you must check whether the comments lead to the song; if they only talk about the bit, the song has not landed.

Criteria for immediately changing the visuals

After plays exceed 2,000, if any of the following occurs, prioritize changing the visuals or the first-screen copy hook, and don't blame the song first:

  • 2-second bounce rate > 50%
  • Average watch time < 4 seconds
  • Like rate < 3%

Criteria for continuing to test the visuals

When most of the following conditions are met, it means users are at least willing to pause for a moment; don't change the visuals immediately:

  • 2-second bounce rate < 42%
  • 5-second completion rate > 22%
  • Like rate > 5%

If saves and comments are still low, prioritize changing the lyric hook, melodic memory point, vocal, comment hook, and release package.

Criteria for changing the song in the second round

If one clear optimization has already been completed, and after the second round exceeds 3,000 plays most of the following conditions still occur, directly change the song or rewrite the hook:

  • Like rate < 4%
  • Save rate < 0.3%
  • Comment rate < 0.2%
  • Share rate < 0.1%
  • 5-second completion rate < 20%
  • Completion rate < 5%

Criteria for continuing to optimize the same direction

When most of the following conditions are met, continue optimizing the same direction; don't change the song immediately:

  • The 2-second bounce rate has not exceeded 50%.
  • The 5-second completion rate reaches or approaches 20%, but is still not strong enough.
  • At least one of likes, saves, comments, or shares is clearly not bad.
  • People in the comments ask for the song name, request the full version, request it be released, ask where to listen, or someone copies out the lyrics.
  • The save rate is higher than the comment rate, indicating the song or visuals have retention value.
  • There is a signal in the share rate, indicating the emotional expression has the ability to speak for users.
  • Original sound usage or follow-along videos begin to appear, even if the original video's plays are not especially exaggerated.

When continuing to optimize, change only one core variable: if the first 2 seconds are weak, only change the visuals / first-screen copy hook; if 2-5 seconds are weak, only change the first line of lyrics / vocal entry / melodic memory point; if interaction is weak, only change the title / pinned comment / comment hook.

Criteria for changing only the title / hook / cover

When the video itself has some carrying power, but interaction or click intent is weak, don't change the song first; prioritize changing only the title, first-screen copy hook, cover, or pinned comment.

Typical signals:

  • The 2-second bounce rate is not high.
  • The 5-second completion rate is not lower than 20%.
  • The completion rate and average watch time are still okay, but comments, saves, and shares are on the low side.
  • The like rate is okay, but the comment section has no direction, and viewers don't know what to say.
  • High saves but low comments: the song has value, but the comment hook is weak.
  • High comments but low saves: the topic is hot, but the song's memorable hook or reason to save is weak; first revise the pinned comment and title, and see whether you can steer comments toward the song.
  • The title only describes facts and does not create a relationship, suspense, identity, or emotional entry point.

The standard for when a song has potential but the video did not carry it

When any of the following combinations is met, do not change the song first; change the video packaging first:

  • The 2-second bounce rate is > 50%, but there are clear positive signals in saves, comments, shares, requests for the song name, or requests for the full version.
  • The save rate or share rate is relatively strong, and in the comments people are copying lyrics, asking for the song name, or saying they want to hear the full version, but completion rate or plays are not ideal.
  • The original creator's single-video plays are not exaggerated, but use of the original sound, duets, or derivative creations are starting to grow.
  • The comment section is talking about the song, for example "what's the song name," "where can I listen," "is it released," "this line hurts," "I want to hear the full version," rather than only praising looks, visuals, jokes, or footage.

How to write the conclusion:

这不是先换歌的情况。更像是歌有潜力，但当前视频没有把歌曲价值在前 5 秒送到观众面前。

Key branching

  • The first round is poor, but the like rate is still okay: prioritize changing the visuals / editing / first-screen copy hook, then test once more.
  • The second round is still poor: do not continue fine-tuning the title; change the song or rewrite the Hook.
  • The first 2 seconds are poor: check the visuals first; do not blame the song first.
  • The 2-5 seconds are poor: check the short lyric line, the position of the melodic hook, vocal entry, and entry into the main section.
  • If the first lyric line is not long and the vocal has also entered, but the 2-5 seconds are still weak, check whether the lyrics lack tension: no specific desire, no obstacle, no disguise, no relational repression of "shouldn't but still," and no unfinished moment.
  • Interaction is poor: check the lyric theme, comment hook, title angle, and target audience.

Output format

数据可用性：样本足够 / 样本不足 / 变量混乱

核心结论：
{一句话}

本轮测试阶段：第一轮 / 第二轮 / 信息不足无法判断

关键指标判断：
- 模板类型：
- 最接近案例：
- 本类型主指标：
- 本类型不应过度解读的指标：
- 曲风类型：
- 最接近曲风卡 / 曲风组：
- 本曲风主指标：
- 本曲风不应过度解读的指标：
- 2 秒：
- 5 秒：
- 完播：
- 点赞：
- 评论：
- 收藏：
- 转发：

问题更可能在：
1.
2.
3.

决策：
- 是否立刻换画面：
- 是否继续测当前歌曲：
- 是否直接换歌 / 重写 Hook：
- 是否只换标题 / 钩子 / 封面：
- 是否属于“歌有潜力但视频没承接住”：

下一步动作：
1.
2.
3.

2-5 秒专项判断：
- 画面承接：
- 第一句歌词长度：
- 旋律记忆点位置：
- 是否需要短句化 / 记忆点前置：

是否值得继续测试：是 / 否 / 需要补样本

是否值得扩写完整版：是 / 否 / 暂时不能判断

下一步推荐：/bmz-xxx
原因：{一句话说明为什么下一步应该进入这个模块}
你可以直接回复：{用户下一条最省事的回复方式}

是否需要用 /bmz-case 沉淀目前这套模板？

Mandatory rule for retaining cases after review

Every time after the user sends you data for review, as long as you have completed the data review and given next-step recommendations, at the end, in addition to the normal "next-step recommendation," you must ask one more question:

是否需要用 /bmz-case 沉淀目前这套模板？

This sentence does not replace the "next-step recommendation"; it is appended in addition. Even if you recommend that the user next go to /bmz-lyrics、/bmz-preflight、/bmz-publish、/bmz-prompt、/bmz-positioning or continue /bmz-review, you must still ask.

If this round of data is clearly positive, it means it can be retained as a successful template. If this round of data is clearly negative, it can also be retained as avoidance experience or a failure variable. If this round of data is moderate but has potential, it can be retained as a template pending retesting.

Evidence level

When outputting judgments, recommendations, routes, scores, reviews, or SOPs, you must mark the evidence level as much as possible:

  • S level: verified by real publishing backend data. Comes from real publishing data and reviews from users, teams, users, or bmzskill's own cases, and can be either positive verification or negative verification.
  • A level: public market benchmarking. Comes from reviewable public videos, public accounts, public rankings, public comment sections, public platform data, or public videos/screenshots provided by users.
  • B level: industry practical methods / bmzskill methodology. Comes from methods repeatedly used in copyright companies, music promotion, short-video operations, and creative practice, or methods currently organized by bmzskill that are pending case calibration.

If there is no S/A level evidence, do not state recommendations as absolute conclusions. You should write clearly:

证据等级：B 级
依据：{为什么这样判断}
不确定性：{缺少什么数据或对标}
验证方式：{下一步如何用短视频数据验证}

V1 uses only the three levels S / A / B and does not expand to more levels.