Stacks

Somebody’s working setup, with the reasoning attached and a name on it. A set is things that go together. A flow is the same picking read in order: this is how I work on a monolith, and here is what I run at each point. The picking is the author’s. The numbers on each item are read off its repository and look the same wherever that item turns up, including inside somebody else’s stack.

• Write the thing that does not exist yet

  For the point where the answer stops being "install something" and starts being "build it". Two Anthropic meta-skills, read together rather than in sequence, because the skill-or-server decision is what the pair of them makes legible and neither one makes on its own.

  1. 1Skill creator
  2. 2Mcp builder

  agentscli2 steps

• Write against the API that exists today

  The worst failure in third-party integration work does not look like a failure. The call is idiomatic, the types line up, and the method was removed two majors ago. Two sources of truth on two different clocks, and the case that matters is not when they disagree but when one of them is silent.

  • Context7
  • GitHub

  agentscli2 items

• Read a codebase you did not write

  Three servers that answer the three different questions an agent gets wrong on a repository it has never seen: where is anything, why is it like that, and what version am I calling. Curated stacks in the wild carry all three.

  • Serena MCP: the IDE for your agent
  • GitHub
  • Context7

  agentscli3 items

• Find the cause before writing the fix

  A stack trace tells you where the program died, not where it went wrong. Two servers that answer the two questions the trace cannot: what actually happened out there, and why the code is shaped like that. This is the pairing people doing incident work are documented as running.

  1. 1Sentry mcp
  2. 2GitHub

  agentscli2 steps

• Build the design, then go and look

  An agent implementing a design generates pixels and reasons in text. It will tell you the header is blue and centred while it renders red and left. Read the design as data, then render it in a real browser and compare. This loop is the one the write-ups actually describe.

  1. 1Figma Context MCP
  2. 2Playwright mcp

  agentscli2 steps

• A review that is not a rubber stamp

  Three systems that can each contradict the pull request description in a different way. This is the reach layer, it is the part of the problem the ecosystem has actually solved, and on its own it is not yet a review. The page says which half is missing.

  1. 1GitHub
  2. 2Context7
  3. 3Sentry mcp

  agentscli3 steps