AgentSkillsCN

10-10-post

针对草稿进行迭代评分循环。依据5项核心指标逐一评估，重点改进薄弱环节，反复打分直至各项指标均达到8分以上，最多迭代5次。 触发条件：“为这篇帖子打分”“满分10分”“评价这份草稿”“这稿子够好吗？”

SKILL.md
--- frontmatter
name: 10-10-post
description: |
  Iterative scoring loop for post drafts. Evaluate on 5 criteria, fix weak
  areas, re-score until all criteria hit 8+. Max 5 iterations.
  Triggers: "score this post", "10/10", "rate this draft", "is this good enough".
user-invocable: true

10/10 Post

Iterate on a post draft using structured evaluation until it scores 8+ on every criterion.

The Loop

Step 1: Evaluate

Score the draft on 5 criteria (each 1-10):

CriterionWhat to Assess8+ Means
DepthIs there a structural insight beyond the surface story? Does "why" go 2+ levels deep? Is there a broader principle?Reader learns something about engineering, not just about what happened
SpecificityReal numbers from real code? Exact values, not rounded? Details that only come from reading the source?Post could only be written by someone with access to this codebase
Non-obviousWould this surprise someone who builds similar systems? Is there a "why behind the why"?Reader's reaction is "I wouldn't have thought of that"
Human voiceContractions present (5+)? Sentence lengths varied? Vulnerable moment? Parenthetical aside? Sounds like a person?You'd believe a human typed this on their phone
Anti-AIZero em dashes? Zero AI vocabulary? No negative parallelisms? No synonym cycling? No closing question? Varied structure?A skeptic scanning for AI tells finds nothing

Step 2: Identify Weaknesses

For each criterion scoring below 8:

  • •Name the specific element that's weak
  • •Quote the exact text that needs fixing
  • •Describe what "8+" looks like for this specific post

Step 3: Fix

Make targeted fixes for each weakness:

  • •Depth below 8: Add the second "why." Add the structural principle. Connect to a broader engineering lesson.
  • •Specificity below 8: Go back to the source file. Find the exact number, the exact config value, the exact timing. Replace vague language with it.
  • •Non-obvious below 8: Ask: "What would someone assume about this situation? How is reality different?" Add that contrast.
  • •Human voice below 8: Add contractions. Break a long sentence into a short burst. Add a parenthetical aside. Add a moment of honesty.
  • •Anti-AI below 8: Scan for em dashes, AI vocabulary, negative parallelisms, synonym cycling. Fix every instance found.

Make 1-3 targeted changes per criterion. Don't overhaul everything at once.

Step 4: Re-evaluate

Go back to Step 1. Score the fixed draft on all 5 criteria.

Stop Condition

Stop when ALL 5 criteria score 8+.

Escape Hatch

If after 5 iterations a criterion is stuck below 8:

  • •Flag which criterion is stuck and why
  • •Present the best version to the user with the limitation noted
  • •Suggest what additional source material might help (e.g., "if we had the actual latency numbers from the monitoring dashboard, Specificity would jump to 9")

Output Format

Per Iteration

code
=== Iteration [N] ===

| Criterion | Score | Notes |
|-----------|-------|-------|
| Depth | X/10 | [specific assessment] |
| Specificity | X/10 | [specific assessment] |
| Non-obvious | X/10 | [specific assessment] |
| Human voice | X/10 | [specific assessment] |
| Anti-AI | X/10 | [specific assessment] |

Weaknesses: [list what needs fixing]
Fixes applied: [list what was changed]

Final

code
=== 10/10 Post Result ===
Iterations: [N]
Final scores: Depth [X] | Specificity [X] | Non-obvious [X] | Human [X] | Anti-AI [X]
Status: ALL 8+ | STUCK ON [criterion]

Rules

  • •Be honest in scoring. Inflated scores defeat the purpose.
  • •A 7 is not an 8. Close doesn't count.
  • •Score against the rules in .claude/rules/, not against your own preferences.
  • •One fix at a time per criterion. Don't rewrite the entire post each iteration.
  • •Max 5 iterations. If not 10/10 by then, ship the best version with notes.