菜单
Estimate software performance, capacity, latency, throughput, and cloud cost from first-principles rough numbers. Use when the user says something is too slow, asks whether a design will scale, gives stack numbers, or wants early tradeoffs before benchmarking.
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
Create a researched, durable, approval-gated goal contract whose verification tools, repository-vault authentication, exact proof steps, rollout safety, and evidence lifecycle are ready before execution. Use when the user asks to set, create, refine, save, or verify a goal; asks whether an agent has everything needed to prove completion; wants a proof-driven alternative to a raw `/goal`; or explicitly invokes `$verify-goal`.
Relentlessly interview the user in dependency-aware batches to sharpen a plan, decision, design, idea, codebase change, or bug fix. Use when the user wants to get grilled, investigate a bug from a ticket or screenshots, pass unfinished grilling to a teammate, or pick it up from a PR. Begin each session by choosing native dialogs or one full text batch, then Sideshow or experimental Grill Visuals for live diagrams. In repositories, inspect current behavior, run a mandatory gap and blast-radius pass, and maintain durable decision docs as answers settle.
Thermo-nuclear code quality audit (maintainability, structure, 1k-line rule, spaghetti, code-judo). Invoked via Task after a parent gathers diff and file contents. Loads the rubric from the `thermo-nuclear-code-quality-review` skill in the cursor-team-kit plugin.
Validate your code changes through the no-mistakes pipeline - automated code review, tests, lint, docs, push, PR, and CI - before they reach upstream. Use when the user asks to run no-mistakes, gate or ship or validate their changes, push safely, asks you to do a task and then validate it, or invokes /no-mistakes.
Ask which skill or flow fits your situation. A router over the user-invoked skills in this repo.
Scan a codebase for deepening opportunities, present them as a visual HTML report, then grill through whichever one you pick.
| name | napkin-math |
| description | Estimate software performance, capacity, latency, throughput, and cloud cost from first-principles rough numbers. Use when the user says something is too slow, asks whether a design will scale, gives stack numbers, or wants early tradeoffs before benchmarking. |
Use rough orders of magnitude to find the limiting resource before coding. Bias to useful bounds, not faux precision.
Name the path - turn the request into one concrete operation: request, job, sync, import, query, deploy, render, or checkout. Done when the hot path is a short ordered list.
Attach units - write input assumptions in visible units: RPS, rows, bytes, objects, users, fanout, regions, retries, retention, concurrency. Done when every variable has a unit and a rough value.
Load current numbers - when any estimate depends on hardware, network, storage, serialization, compression, or cloud cost, read sirupsen/napkin-math, especially its README tables, unless repo-local production measurements are more relevant.
Done when the numbers used are either copied from the repo-local system or traced to the napkin-math reference.
Budget each step - estimate latency, throughput, and cost for each step using the numbers below, napkin-math, or repo-local measurements. Done when each step has an optimistic and pessimistic bound.
Find the wall - compare required work against the slowest step or tightest quota. Done when one bottleneck is named with math, not vibes.
Trade one constraint - propose 2-4 changes that move the wall: batch, cache, parallelize, colocate, compress, denormalize, drop durability, precompute, or reduce fanout. Done when each tradeoff says what improves, what worsens, and what must be measured next.
Round these hard. Prefer powers of ten when unsure.
| Operation | Rough budget |
|---|---|
| CPU cache / sequential memory | ns to tens of ns |
| System call | hundreds of ns |
| SSD sequential read | ~GB/s |
| SSD random read | ~100 us |
| Context switch | ~10 us |
| Same-zone network | ~100 us, many GB/s |
| Same-region network | few hundred us |
| Cross-region network | tens to hundreds ms |
| Redis/MySQL/cache query | ~0.5 ms before app logic |
| Blob GET/PUT first byte | tens to hundreds ms |
| JSON-ish serialization | ~100 MB/s |
| Efficient binary serialization | ~1 GB/s |
| Compression/decompression | hundreds MB/s to ~1 GB/s |
| Internet egress | expensive enough to check |
| Logs/traces | often costlier than expected |
These are fallback memory anchors. For real estimates, use provider docs, production traces, or sirupsen/napkin-math when a row matters.
Return:
## Assumptions
- ...
## Back-of-envelope
| Step | Work | Budget | Result |
| --- | --- | --- | --- |
## Bottleneck
...
## Tradeoffs
- ...
## Measure Next
- ...