R
robertmnyborg/analyst-rigor-skills
Claude Code skills that stop an agent from fabricating analytical conclusions: causal claims, metric…
更新 2026-11-18
在 GitHub 查看 ↗直链 · 无中转页
装完之后怎么算装好了
- 装完该看到什么 — 命令正常结束时退出码为 0,终端里没有 "command not found" 或 "permission denied"。
- 一行自检 — 按该仓库 README 里给出的最小用法跑一次;能按 README 的描述输出就是通了。
- 答不出来怎么办 — 这个项目没有写排错说明,去仓库的 Issues 搜 install。
安装命令取自该仓库;以上判据只描述可观察到的结果,不替代项目自己的文档。
- ★ 数取值
- 0 ★ · robertmnyborg/analyst-rigor-skills · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-19 · 最近见到 2026-11-20 · 共见到 2 次 · 被 2 个来源列出
- 评分构成
- 综合分 203 · 2 个来源 · repo
- 安装
- 安装命令见上方「安装与配置」。
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: Agent 技能。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
robertmnyborg/analyst-rigor-skills;我们收录于 2026-11-19;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。README 原文(英文,未翻译) · 在 GitHub 读完整 README →
# analyst-rigor-skills
以下摘录自该仓库的 README(片段)
Five Claude Code skills that stop an agent from fabricating analytical conclusions: a confounded group comparison reported as an effect, a same-day coincidence named as root cause, an invented customer quote, a persona with no buyer behind it, a feature list passed off as a strategy.
Each rule exists because an agent shipped the wrong answer without it.
## Install
Claude Code:
/plugin marketplace add robertmnyborg/claude-harness
/plugin install analyst-rigor-skills@claude-harness
Codex, Cursor, and other agents that read SKILL.md, via the skills CLI:
bash
npx skills add robertmnyborg/analyst-rigor-skills
job-card also calls the job-card-gate script, and the optional hook routes causal questions to causal-claim. For those, install by hand:
### Manual install
bash
git clone https://github.com/robertmnyborg/analyst-rigor-skills
cp -R analyst-rigor-skills/skills/ ~/.claude/skills/
cp analyst-rigor-skills/bin/job-card-gate ~/.local/bin/ # any directory on your PATH
cp analyst-rigor-skills/hooks/causal-claim-router.sh ~/.claude/hooks/
Register the hook in ~/.claude/settings.json (optional, needs jq):
json
"hooks": {
"PreToolUse": [
"matcher": "Bash",
"hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/causal-claim-router.sh" }]
Requirements: Claude Code, Python 3 (gate), bash and jq (hook). Run bash tests/run.sh to check the gate and the hook.
## causal-claim
Failure it prevents: a between-group number (adopters vs non-adopters, before vs after, high vs low tier) reported as an effect while the confounds stay unstated.
Why: in a multi-day outcome investigation, several course corrections were confounds a pre-registration would have caught: composition, exposure window, denominator, data-source mix, and a skipped design ladder.
How: print a 12-row pre-registration table (mechanism, denominator, treatment timing, exposure, falsifier, declared slice count) before the first query, climb a five-rung design ladder from within-entity down to cross-sectional, report against the falsifier, then red-team three objections. Hard stops forbid rung-5 "effects", unstated denominators, and same-day suspects that carried under half the lost volume.
Trigger: hooks/causal-claim-router.sh fires once per session when a Bash command runs SQL that aggregates by a cohort, tier, or treatment column. It adds context and never blocks. It over-fires by design: a false fire costs one nudge, a miss ships an uncontrolled number to an exec slide.
## metric-diagnosis
Failure it prevents: "the metric dropped because of X" where X happened on the same day and nothing else connects them.
Why: in one metric-drop investigation, a draft blamed a change made the same day that touched only a small share of the lost volume, and a later draft rested on a "clicks" series that mostly counted an internal API call. The skill had already fired; both wrong answers shipped anyway, so the gates below are the fix.
How: clarify and row-sample every series, size the move against noise and YoY, decompose with a metric tree, segment (TROPICS, provider first), rank hypotheses, then pass three gates before writing "root cause": delta attribution (the suspect carried at least half the lost volume), observed mechanism, and run the confirming query yourself. Works in live-data mode and in PM/DS interview mode.
Each rule exists because an agent shipped the wrong answer without it.
## Install
Claude Code:
/plugin marketplace add robertmnyborg/claude-harness
/plugin install analyst-rigor-skills@claude-harness
Codex, Cursor, and other agents that read SKILL.md, via the skills CLI:
bash
npx skills add robertmnyborg/analyst-rigor-skills
job-card also calls the job-card-gate script, and the optional hook routes causal questions to causal-claim. For those, install by hand:
### Manual install
bash
git clone https://github.com/robertmnyborg/analyst-rigor-skills
cp -R analyst-rigor-skills/skills/ ~/.claude/skills/
cp analyst-rigor-skills/bin/job-card-gate ~/.local/bin/ # any directory on your PATH
cp analyst-rigor-skills/hooks/causal-claim-router.sh ~/.claude/hooks/
Register the hook in ~/.claude/settings.json (optional, needs jq):
json
"hooks": {
"PreToolUse": [
"matcher": "Bash",
"hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/causal-claim-router.sh" }]
Requirements: Claude Code, Python 3 (gate), bash and jq (hook). Run bash tests/run.sh to check the gate and the hook.
## causal-claim
Failure it prevents: a between-group number (adopters vs non-adopters, before vs after, high vs low tier) reported as an effect while the confounds stay unstated.
Why: in a multi-day outcome investigation, several course corrections were confounds a pre-registration would have caught: composition, exposure window, denominator, data-source mix, and a skipped design ladder.
How: print a 12-row pre-registration table (mechanism, denominator, treatment timing, exposure, falsifier, declared slice count) before the first query, climb a five-rung design ladder from within-entity down to cross-sectional, report against the falsifier, then red-team three objections. Hard stops forbid rung-5 "effects", unstated denominators, and same-day suspects that carried under half the lost volume.
Trigger: hooks/causal-claim-router.sh fires once per session when a Bash command runs SQL that aggregates by a cohort, tier, or treatment column. It adds context and never blocks. It over-fires by design: a false fire costs one nudge, a miss ships an uncontrolled number to an exec slide.
## metric-diagnosis
Failure it prevents: "the metric dropped because of X" where X happened on the same day and nothing else connects them.
Why: in one metric-drop investigation, a draft blamed a change made the same day that touched only a small share of the lost volume, and a later draft rested on a "clicks" series that mostly counted an internal API call. The skill had already fired; both wrong answers shipped anyway, so the gates below are the fix.
How: clarify and row-sample every series, size the move against noise and YoY, decompose with a metric tree, segment (TROPICS, provider first), rank hypotheses, then pass three gates before writing "root cause": delta attribution (the suspect carried at least half the lost volume), observed mechanism, and run the confirming query yourself. Works in live-data mode and in PM/DS interview mode.
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/robertmnyborg-analyst-rigor-skills/)
- 来源
- github · https://github.com/robertmnyborg/analyst-rigor-skills
- 按领域浏览
- Agent 技能
- 收录来源
- 2 个独立来源
- 收录于
- 今天
更多条目
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。