Help帮助

01Getting started上手

Press ⌥Space anywhere to bring Napkin up. Type one thing per line; the answer appears on the right as you type. Press Esc to put it away — focus returns to whatever you were doing. 在任何地方按 ⌥Space 呼出 Napkin。 一行写一件事，右边随打随出结果。按 Esc 收起， 焦点回到你刚才那个窗口。

A line that isn't a calculation is just a note — write whatever you want between the numbers. Incomplete lines stay silent instead of flashing an error at you mid-keystroke. 不是算式的行就当注解，数字之间想写什么写什么。 没写完的行会保持安静，不会打到一半就红给你看。

02Numbers & math数字与运算

Functions: sqrt abs round floor ceil min max sin cos tan log ln log2 exp 函数：sqrt abs round floor ceil min max sin cos tan log ln log2 exp

That last one is not a typo. Napkin does money in fixed-point decimal, so 0.1 + 0.2 is 0.3 — not 0.30000000000000004. 最后一行不是笔误。Napkin 用定点十进制算钱， 0.1 + 0.2 就是 0.3，不是 0.30000000000000004。

03Units单位

Write the unit next to the number and Napkin carries it through the whole calculation. This is the part that catches mistakes: units that don't match refuse to combine instead of quietly producing a number. 把单位写在数字旁边，Napkin 会让它跟着整个算式走。 拦错就靠这个：单位对不上时它拒绝计算，而不是安静地给你一个数。

Converting换算

Use in, to, as or into. 用 换、换算成、转成，英文写 in / to。

Exchange rates are fetched daily and cached. Everything else (length, mass, data, time) is an exact definition and works offline. 汇率每天取一次并缓存。其他的（长度、质量、数据、时间） 都是精确定义，离线一样算。

Compound units复合单位

Divide units to build rates, and they cancel the way they should. Napkin cancels by dimension, not by name — so users against user still cancels. 用除号构造速率，该约掉的会约掉。 Napkin 按量纲约分而不是按名字，所以 users 和 user 也能对上。

When units don't match单位对不上时

What it knows认得哪些单位

04Percentages百分比

"Up 12%" and "up 0.61 points" are different claims, and mixing them up is how a deck ends up wrong. Napkin keeps them apart. 「涨了 12%」和「涨了 0.61 个点」是两回事，混着说汇报就错了。 Napkin 把这两个分得很清楚。

+ 15% means relative (a 15% increase). + 0.61pp means absolute (add 0.61 percentage points). Use pp whenever you mean points. + 15% 是相对变化（涨 15%）。 + 0.61pp 是绝对变化（加 0.61 个百分点）。 说「点」的时候一律写 pp。

05Variables变量

Name your assumptions instead of scattering raw numbers. Change one and everything below recomputes. Both = and : work; names can be in any language and can contain spaces (monthly cost, AI bill). 给假设起名字，别把裸数字撒得到处都是。改一个，下面全部重算。 = 和 : 都行，名字可以用中文，也可以带空格 （monthly cost、AI bill）。

Select a few lines and the total appears at the bottom — no need to write sum at all. There is deliberately no whole-document total: most drafts have several unrelated calculations in them, and summing all of them produces a meaningless number. 选中几行，底部直接出合计，连 sum 都不用写。 刻意不做全文档合计——一张草稿上通常有好几段互不相干的账， 全加起来是个没有意义的数。

06Dates & time日期与时间

Month and year arithmetic is calendar-aware: adding one month to Jan 31 lands on Feb 28, not Mar 3. Time zones work too — now in PST. 月和年的加减按日历走：1 月 31 日加一个月落在 2 月 28 日， 而不是 3 月 3 日。时区也认，写 now in PST 或 北京时间。

07A/B testingA/B 实验

Give it the two arms and it runs a two-proportion z-test: the lift, the p-value, and a straight verdict on whether the difference could just be noise. 给它两组数据，它跑双比例 z 检验：提升幅度、p 值， 以及一个直接的结论——这个差异是不是可能纯属噪声。

Arguments are exposures A, conversions A, exposures B, conversions B. The A: line stays silent by design — the comparison only exists once there's a second arm. 参数顺序是 曝光A, 转化A, 曝光B, 转化B。 A: 那行故意不出结果——有了第二组才谈得上比较。

Planning a test开跑之前

samplesize answers "how many do I need?"; mde answers the reverse — "with the traffic I have, how big does the effect need to be before I could detect it?" Both assume 80% power and α = 0.05. If you divide by a literal rate rather than a variable, parenthesize it: / (12k users/day) — otherwise day reads as a second division. 样本量 回答「我需要多少人」；mde 回答反过来的问题 ——「就这么点流量，效果得多大我才看得出来」。两者都按 80% 功效、α = 0.05 算。 除数如果直接写成带单位的字面量，要加括号：/ (12k users/day)—— 不加的话 day 会被当成又一次除法。用变量名就没这个问题。

Two-tailed only. One-tailed tests halve the p-value and are the most common way an experiment gets talked into looking significant. Napkin doesn't offer them. 只做双尾。单尾把 p 值砍一半，是把实验「说成」显著的最常见入口。 Napkin 不提供这个选项。

08Uncertainty区间估算

You rarely know ARPU exactly. Write the range with ± (or +-) and Napkin carries it all the way down, so the final number admits how rough it is. 客单价你从来就不知道确切值。用 ±（或 +-） 写下浮动范围，Napkin 会一路带到底，最后那个数会老实告诉你它有多粗。

Errors combine as root-sum-square, not by adding up. Adding them linearly assumes every assumption is wrong in the same direction at once, which explodes the range as soon as you have more than two or three variables. 误差按平方和开根合成，而不是直接相加。 线性相加等于假设所有假设同时朝一个方向错， 变量一多，区间就炸得没法看了。

Write the unit on the number, before the ±. 39 $/user ± 15% works; 39 ± 15% $/user does not parse the way you'd expect. 单位要写在数字上，写在 ± 前面。 39 ¥/user ± 15% 可以；39 ± 15% ¥/user 解析结果和你想的不一样。

09Shortcuts快捷键

10Known limits已知限制

Things Napkin currently gets wrong or doesn't do. Listed here rather than left for you to discover mid-calculation. Napkin 目前做不到或者会做错的地方。写在这里， 省得你算到一半才发现。

Found something else? Open an issue. 发现别的问题？提个 issue。