菜单
Comprehensive peer review from skeptical economist and "lazy AI" perspectives. Identifies missing LaTeX equations, charts, hardcoded values, weak arguments, and context-loss vulnerabilities. Reviews calculation chains for all parameters.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Critique book prose in the author's voice using patterns learned from his own git edit history (not a hand-written rubric). Use when asked to "voice-check", "punch up", "make this sound like me / like Vonnegut / like Cunk", or to review whether a passage is too pompous / cute / self-aware. Catches what the regex scanner (voice-check.py, ~1.4% of real edits) misses.
Review and improve QMD chapter writing quality by replacing boring/consultant text with Wishonia's deadpan alien voice, consolidating redundancy, and reducing word count. Use when asked to fix voice, make content funnier, remove jargon, or enforce the book's writing style.
Spawns parallel agents to review all paperback QMD files. Reads variable-resolved previews, flags judgment-required issues with REVIEW comments.
Automated book maintenance. Runs health checks, prioritizes work, and systematically fixes issues. Use when you want Codex to autonomously improve the book.
Check GitHub Actions workflow status, diagnose failures, fetch logs, and fix common build issues. Use when builds fail, deploys break, or you need to investigate CI/CD problems.
Critically analyze a calculated parameter for academic rigor. Checks methodology, identifies weaknesses, suggests improvements for economist audiences.
| name | peer-review |
| description | Comprehensive peer review from skeptical economist and "lazy AI" perspectives. Identifies missing LaTeX equations, charts, hardcoded values, weak arguments, and context-loss vulnerabilities. Reviews calculation chains for all parameters. |
| allowed-tools | ["Read","Edit","Grep","Glob","Bash","Write","TodoWrite"] |
Four-perspective review AND fix: (1) Skeptical economist, (2) Lazy AI vulnerabilities, (3) Calculation audit, (4) Enhancement opportunities. Generates report then implements all high/medium priority fixes. If no file specified, ask which to review.
Read the file, then:
grep "_latex\":" _variables.yml | head -100
ls knowledge/figures/*.qmd | head -50
Read as a hostile peer reviewer skeptical of large claims.
| Category | Check |
|---|---|
| Methodology | Causal claims justified? Counterfactual specified? CIs appropriate? Assumptions stated? |
| Sources | Peer-reviewed? Old (>5y)? Over-reliance on single studies? |
| Logic | Conclusions follow? Unstated assumptions? Strongest counterargument addressed? |
| Framing | Neutral language? Limitations discussed? Costs AND benefits quantified? |
| Pattern | Example | Fix |
|---|---|---|
| Problem-solution mismatch | Defending against Goodhart when metrics are real-world outcomes (income, mortality) that can't be gamed without improvement | Ask: does this problem actually apply here? Cut irrelevant defenses |
| Speculative mechanisms | "staking," "immutable rules," "distributed ledgers," "zero-knowledge proofs" when simpler solutions exist | Replace with realistic mechanisms: institutional diversity, median aggregation, transparency |
| Internal contradictions | "Immutable rules solve X" then later "immutable rules are insufficient" | Grep for key claims, verify consistency throughout |
| Overconfident claims | "trivial to influence," "guaranteed," "impossible" | Soften: "incentive is enormous," "difficult," "unlikely" |
| Self-congratulatory framing | "A sophisticated objection," "The honest assessment:" | Cut framing, state content directly |
| Conflated problems | Treating Goodhart gaming (behavior) and measurement capture (methodology) as same problem | Distinguish clearly; different problems need different solutions |
| Over-defense | 60 lines defending against non-problem, key point buried at end | Lead with "why this doesn't apply," cut irrelevant mitigations |
| Unexplained jargon | "mētis," "Schelling points," "Sybil attacks" | Define inline or replace with plain language |
| Academic pomposity | "This paper makes three contributions," "We formalize the X Condition" | Cut if content is covered elsewhere; keep only if adds navigation value |
Key question: For each defense/mitigation, ask "What specific attack does this prevent?" If answer is vague or doesn't apply to actual proposal, cut it.
Write simply for general audience without losing precision. Target 8th-grade reading level, 10-15 word sentences average. Can a smart 8th-grader understand this? If not, simplify. Never sacrifice accuracy—use precise terms but define immediately. Break complex ideas into 2-3 short sentences. Delete ruthlessly.
| Delete/Replace | Example |
|---|---|
| Redundancy | "past history" -> "history" |
| Hedging/blather | "It could potentially be argued that" -> delete or "This suggests" |
| Filler phrases | "It is important to note that", "In order to", "the fact that" |
| Nominalizations | "the implementation of" -> "implementing" |
| Passive voice | "was conducted by" -> "X conducted" |
| Em-dashes (—) | Replace with period (preferred), comma, parenthesis, or semicolon |
| Long sentences | "X happened, which caused Y, and this led to Z" -> "X happened. This caused Y. Y led to Z." |
Flag and fix: Sentences >25 words, paragraphs >5 sentences, sections repeating earlier content, jargon without definition.
Inject deadpan, absurdist humor when it clarifies without detracting from credibility.
| Use when | Avoid when |
|---|---|
| Highlighting absurd contradictions/inefficiencies | Methodology sections (maintains rigor) |
| Making abstract concepts concrete via absurd comparisons | Serious topics (deaths, suffering, failures) |
| Breaking up dense technical sections | Executive summaries/abstracts |
| Calling out obvious but unstated truths | If it undermines credibility |
Style guide:
Examples: ❌ "Regulatory delays kill people. LOL." ✅ "Regulatory delays kill people. This is not controversial. Yet the system continues."
What survives if AI only reads first 3 paragraphs?
| Pattern | Fix |
|---|---|
| Late Caveats | Front-load qualifications |
| CI Collapse ("102M" loses CI) | Add CI inline at every mention |
| Scope Loss | Repeat scope with each figure |
| Conditional Loss | Add "(under X assumption)" inline |
| Conservative framing loss | Bold the conservative note |
python scripts/preview-qmd-with-variables.py <file> --numbers-only
Red flags: Dollar amounts, percentages, large numbers not using {{< var ... >}}. Numbers in $$ blocks should use _latex variables.
For calculated values, check for _latex version: {{< var parameter_name_latex >}}
Prefer @citation-key from references.bib over manual superscripts. Web search for missing sources.
Check knowledge/figures/*<variable>*.qmd for tornado, mc-distribution, sensitivity-table charts.
For each calculated parameter:
grep -B5 -A50 "^PARAMETER_NAME = Parameter" dih_models/parameters.py
| Check | Criteria |
|---|---|
| Distribution | Normal (symmetric), Lognormal (costs, positive), Beta (probabilities), Fixed (constants only) |
| Formula | No double-counting, independent factors, units match |
| CI Width | <±20% overconfident, ±30-100% reasonable, >±200% uninformative |
| Sources | Meta-analysis > peer-reviewed > government stats > industry reports > assumptions (flag prominently) |
## Peer Review: <filename>
### Summary
[1-2 sentences]
### Issues Found
| Issue | Type | Severity | Location | Fix |
Types: methodology, sources, logic, framing, clarity, lazy-ai, hardcoded, calculation, enhancement
### Priority Actions
1. **High:** 2. **Medium:** 3. **Low:**
Do not just report. Fix all High/Medium issues using Edit tool.
Re-read edited sections to verify. Track remaining issues with TodoWrite.