OPC Short Video Review
Purpose
Use this skill as a closed loop: pre-publish script review -> publish hypothesis -> post-publish data review -> reusable experience. Treat 爆款潜力 as a directional, explainable judgment, not a promise. Every recommendation must say what to change, why it should affect a metric, and how to verify it after publishing.
Runtime Portability
Keep the skill portable across Codex, Claude Code, OpenClaw, Hermes, and similar agents:
- Use only
nameanddescriptionin required frontmatter. Do not rely on platform-specific metadata to trigger the core workflow. - Resolve supporting files relative to the folder containing this
SKILL.md. If the runtime cannot access supporting files, answer from this file and ask the user to provide the relevant reference text. - Before using shell scripts, confirm the runtime has a local shell and the uploaded file has a local path. If not, parse tables/screenshots manually.
- Do not auto-write case-library files or make git commits. Return proposed entries first; only persist them when the user explicitly asks.
- For install paths and packaging notes, read runtime-portability.md.
Mode Selection
Choose the workflow from the user's input:
- Script Audit: draft script, outline, title, cover, storyboard, or multiple variants.
- Variant Comparison: multiple hooks, covers, titles, openings, scripts, or CTAs.
- Post-Publish Data Review: analytics screenshot, CSV export, table, or reported metrics.
- Cross-Platform Adaptation: adapting one script between Douyin/TikTok/Reels/Shorts/Xiaohongshu/Bilibili/Video Channels/Kuaishou.
- Experience Accumulation: summarizing lessons, updating account-specific calibration, or comparing many past videos.
If the stage is genuinely unclear, ask whether the user wants pre-publish review, variant comparison, or post-publish review. Otherwise proceed with explicit assumptions.
Context Checklist
Gather only what changes the recommendation:
- Platform, account/niche, target audience, account stage, product/offer, desired goal, target duration, script/shot list, cover/title/caption, production constraints.
- For data review: views, impressions/reach, duration, average watch time, retention, completion rate, likes, comments, shares, saves, follows, profile clicks, link clicks, conversions, traffic source, publish time, cover/title, and original script.
- If account history exists, read case-library/calibration-notes.md before scoring. If the file is absent or empty, say the judgment is based on general priors.
Script Audit
Use this workflow for new scripts and variants:
- Read content-platform-profiles.md for the target content platform.
- Read script-audit-rubric.md.
- If calibration notes exist for the account/niche, use them to adjust the generic rubric.
- Return the report using review-report-template.md.
- Include at least one direct rewrite. For weak scripts, prioritize the first 3 seconds and the core promise before polishing style.
- End with a publish hypothesis: the metric that should move if the recommendation is right.
Do not give fake precision such as "爆款率 67%". Prefer a potential band, confidence level, and biggest uncertainty. A numeric score is allowed only as a rough internal scorecard.
Variant Comparison
Use this workflow when the user provides multiple hooks, titles, covers, scripts, endings, or CTAs:
- Read variant-testing.md.
- Compare variants side by side against the user's platform and goal.
- Give a clear winner or a fused version; do not average everything into a bland compromise.
- Design a measurable test with the primary metric, sample rule, and what decision to make afterward.
Data Review
Use this workflow after publishing:
- Read data-review-playbook.md.
- If a CSV path is available and shell execution is allowed, run the metrics helper from the skill directory:
python3 scripts/analyze_metrics.py <path-to-csv>
- Normalize raw counts into rates and compare against the account's own baseline first. The helper provides batch medians and initial signals, but final attribution still requires script/platform context.
- Connect metric behavior back to exact script choices: hook, promise, pacing, proof, emotion, interaction design, CTA, or platform packaging.
- Separate script causes from non-script causes such as cover, title, edit, audio, publish time, platform trend, or randomness.
- Return next tests, not just an explanation of what happened.
Experience Loop
When there is a script review, prepare a case entry with status 待回收数据. When data returns, update the same entry with status 已复盘. Use case-entry-template.md and experience-bank.md.
Persistence rules:
- If the user asks only for advice, include the proposed case entry in the answer but do not write files.
- If the user asks to save/update experience, append to case-library/cases.md and update case-library/calibration-notes.md only when the lesson has evidence.
- Mark one-video findings as low-confidence hypotheses.
- Preserve failed tests and contradictions; they are useful calibration data.
Output Principles
- Lead with a verdict and bottleneck.
- Give concrete rewrites, not generic "更抓人".
- Tie every recommendation to a metric: retention, completion, like, comment, share/save, follow, lead, or conversion.
- Optimize for the user's stated goal; a sales video and a pure reach video should not be judged by the same weights.
- Protect creator voice. Sharpen the script without turning every account into the same viral template.
- Never guarantee virality, revenue, distribution, or algorithmic outcomes.
Resource Index
- runtime-portability.md: install and compatibility notes for Codex, Claude Code, OpenClaw, Hermes.
- content-platform-profiles.md: content platform differences and cross-posting rules.
- script-audit-rubric.md: scoring dimensions and rewrite moves.
- variant-testing.md: hook/title/script comparison and A/B test design.
- data-review-playbook.md: metric diagnosis and next-test logic.
- experience-bank.md: evidence-based lesson format.
- review-report-template.md: pre-publish report template.
- case-entry-template.md: case-library entry template.
- analyze_metrics.py: CSV metric normalization helper.
