| name | article-score-retro |
| description | Use when a platform draft already exists and you need pre-publish scoring, blind prediction, or post-publish retrospective for WeChat, Xiaohongshu, X, TikTok, Facebook, Reddit, or dev.to. |
| version | 1.0.0 |
Article Score Retro
Use this skill when a draft is already written and the next step is to evaluate it before publishing or learn from it after publishing.
Use This For
- scoring a draft before publish
- making a blind prediction before publish
- running a T+3 or T+7 retrospective after publish
- comparing expected performance with actual performance
Do Not Use This For
- topic research
- first-pass writing
- platform rewriting
- cover image generation
Inputs
- article or post path
- target platform
- optional goal
- optional actual metrics for retro
Outputs
- structured score summary
- structured blind prediction
- structured retrospective notes
- short list of rubric update suggestions when useful
- persisted experiment ledger files for score, prediction, and retro
Modes
1. Score
Use when the draft is not yet published.
Output should include:
- overall judgment
- 4-6 dimension scores
- top strengths
- top risks
- concrete revision advice
2. Predict
Use when the draft is about to be published.
Output should include:
- expected performance band
- what the post is likely to do well
- what may underperform
- why this prediction is being made
This prediction should be written before results are known.
The prediction must be saved as a file before publish.
3. Retro
Use when results are already available.
Output should include:
- what worked
- what missed
- where the original prediction was right
- where the original prediction was wrong
- what should change next time
Retro must explicitly reference the earlier saved prediction.
If repeated misses or repeated wins appear, hand the findings to platform-rubric-manager.
Ledger Contract
Every scored or published content unit should have a stable content folder.
Recommended structure:
output/content/{slug}/
source.md
wechat.md
x.md
xiaohongshu.md
tiktok.md
facebook.md
reddit.md
devto.md
score-{platform}.json
predict-{platform}.json
retro-{platform}-t3.json
retro-{platform}-t7.json
Minimum rule set:
score writes a structured score file.
predict writes a blind prediction file before publish.
retro reads the original prediction file before writing conclusions.
- Do not overwrite old prediction files after results are known.
File Expectations
Score File
Should include at minimum:
- platform
- date
- overall judgment
- dimension scores
- strengths
- risks
- revision advice
Prediction File
Should include at minimum:
- platform
- date
- expected performance band
- why this outcome is expected
- main upside
- main risk
Retro File
Should include at minimum:
- platform
- date
- actual result summary
- where the prediction was right
- where the prediction was wrong
- what changes next time
- rubric update suggestion if applicable
Platform Lens
- WeChat: title strength, opening depth, section rhythm, authority, insight.
- Xiaohongshu: cover promise, first-screen hook, saveability, scannability, emotional resonance.
- X: first-line sharpness, point density, repost potential, discussion energy.
- TikTok: first-3-second hook, spoken rhythm, retention potential, simplicity.
- Facebook: readability, social tone, commentability, low-friction sharing.
- Reddit: authenticity, discussion fit, anti-marketing tone, information value.
- dev.to: technical clarity, credibility, example quality, implementation usefulness.
Core Pattern
- Identify the platform first.
- Judge the draft by that platform's native logic.
- Write the score file before publish.
- Write the blind prediction file before publish.
- Do not reuse another platform's rubric blindly.
- Keep prediction separate from retrospective.
- Use the saved prediction during retro.
Common Mistakes
- Scoring all platforms with one universal standard.
- Writing vague praise instead of concrete revision advice.
- Doing retro without recording the original prediction.
- Confusing traffic results with content quality alone.
- Editing a prediction after seeing results.
- Skipping file persistence and leaving the process as chat-only advice.
Quick Rules
- Publish-time judgment and post-publish learning are separate steps.
- Platform fit matters as much as writing quality.
- Short feedback beats generic feedback.
- Retro should improve the next draft, not just explain the last one.
- If it is not written to the ledger, it does not count as part of the workflow.