菜单
Use when comparing an Atlys metric across segments (device, os, geo/country, destination, user type, acquisition) to find where performance is unusually good or bad. Provides normalization for messy data and volume guardrails.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/sidagarwal04/click-a-thon-26-submissions --skill atlys-segment-comparison命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | atlys-segment-comparison |
| description | Use when comparing an Atlys metric across segments (device, os, geo/country, destination, user type, acquisition) to find where performance is unusually good or bad. Provides normalization for messy data and volume guardrails. |
Goal: compare a metric across segments to find where something is unusually good or bad, with enough rigor that a PM can trust it. Compute in ClickHouse; return top-N.
lower(device_type) and os (handle NULL os with android device_type).geoip_country_code (ISO-2), optionally geoip_subdivision_1_code.destination.is_new_user (auth), is_guest, is_enterprise, co_travelers.gclid != '' => paid search.Flat tables:
lower(coalesce(nullIf(device_type,''),'unknown')) AS device_norm,
coalesce(nullIf(os,''), if(lower(device_type)='android','android','unknown')) AS os_norm,
upper(((geoip_country_code,),)) geo
JSON payload tables (same logic wrapping payload.*; see atlys-json-payload-access):
lower(coalesce(nullIf(toString(payload.device_type),''),'unknown')) AS device_norm,
coalesce(nullIf(toString(payload.os),''), if(lower(toString(payload.device_type))='android','android','unknown')) AS os_norm,
upper(coalesce(nullIf(toString(payload.geoip_country_code),''),'??')) AS geo
Always return the denominator so small segments can be discounted.
SELECT
upper(coalesce(nullIf(geoip_country_code,''),'??')) AS geo,
lower(coalesce(nullIf(device_type,''),'unknown')) AS device_norm,
uniqExact(user_id) AS users,
countIf(<success_condition>) AS successes,
round(successes / nullIf(users,0), 4) AS rate
FROM atlys.<table>
WHERE timestamp >= {start:DateTime} AND timestamp < {end:DateTime}
GROUP BY geo, device_norm
HAVING users >= 200 -- ignore tiny segments
ORDER BY rate ASC -- worst-performing first
LIMIT 25;
On a JSON payload table, swap the columns for payload.* (e.g.
uniqExact(payload.user_id), upper(coalesce(nullIf(toString(payload.geoip_country_code),''),'??'))),
add WHERE payload.event = '<stage>' when the success condition is event-type-specific, and
window on payload.timestamp.
Compute the rate for each, the absolute and relative gap, and the sample sizes. Report relative gap as the headline ("iOS converts 15% worse than Android") but only if both segments clear the volume guardrail. If a segment is small, mark confidence Low.
You are not running a full stats test in SQL, but:
users >= 200) before calling a gap real.is_back_filled = 1 and non-empty duplicate_id for behaviour reads.