如何將 Claude 用作軟體架構師：Viktor Skill 指南

It can produce a design, and the design will be greenfield and maximal, because the material it learned from was written by people describing systems that worked at a scale worth writing about. Nobody publishes a post about the boring monolith that is still fine. So you get the architecture of a company larger than yours, on an empty field, with the foreclosures left out. That is correctable, but not by asking better questions about the design - only by changing what the model is allowed to assume.

Because those are the write-ups of organisations that genuinely needed them, and there are no write-ups of the companies that did fine without. The average of the literature is therefore an architecture for a larger company. The cheapest correction is to ask for the same system at one tenth of the load you stated, with the same team, then ask what was removed and at what specific measurable point each removed piece would need to come back.

A list of things you have decided you will not be able to do. Choosing eventual consistency forecloses certain guarantees, choosing a monolith forecloses independent deployment, choosing a particular store forecloses certain shapes of growth. The foreclosures are the decision. A model leads with capabilities every time and will not volunteer what you gave up, so ask for the foreclosure list before the enablement list - and treat an option that appears to have no downsides as suspicious rather than as a recommendation.

Because the boxes are the easy part. A service that does one clear thing is a solved problem. All of the difficulty is in the arrows, and an arrow is a claim that two things can talk which the diagram never justifies. Synchronous or asynchronous, what happens on failure, whether a retry is safe, what the timeout is, whether order matters and what guarantees it, what still works if this arrow is down for an hour. Interrogate the arrows and the design usually changes, often collapsing back into fewer boxes.

Generating the failure space. Asking what are all the ways this can go wrong is a recall problem across an enormous body of postmortems and incident write-ups, which is exactly what these tools are for, and it will surface the failure mode you would otherwise have found in production in eight months. What it cannot do is rank them, because ranking needs your traffic, your team, your tolerance and your money. It generates the list, you order it.

It will give you a number and it has no idea. Throughput, latency and capacity depend on your hardware, your data shape, your query patterns and your access distribution, and the only way to get them is to measure. Treat any performance figure it offers as a placeholder where a load test should go. The same applies to service quotas, managed-database limits, instance sizes and cloud pricing, all of which change and all of which it answers from whatever it last saw.

Because greenfield designs do not need one, and greenfield is what the corpus describes. In reality a target architecture with no route to it is a wish, and the migration is where most of the cost lives. Ask for the path in steps that each ship independently with the system live, and apply one test to every step: could we stop here permanently and still be better off than we are now? If the answer is no, the step is too big. Also ask where the point of no return is.

The download is a ZIP with SKILL.md at the root of the archive rather than inside a nested folder, which is the usual reason an upload fails. In the Claude desktop app open Customize, then Skills, upload the ZIP and toggle it on. Skills need code execution enabled, under Settings then Capabilities. Anthropic's help centre currently lists Skills on Free, Pro, Max, Team and Enterprise, while its Academy tutorial lists Pro, Max, Team and Enterprise, so if you are on the free plan check Settings then Capabilities for your own account rather than trusting either page. In ChatGPT or Gemini there is no upload step, so open SKILL.md, copy the contents and paste them into custom instructions.