Write the thing that does not exist yet

The failure this is for

Somebody writes a SKILL.md, reads it back, ships it, and it never runs. Not errors: never runs. The model reads the description, judges it irrelevant to how the request was phrased, and moves on. There is no log line for a skill that was not consulted.

The MCP version is the same shape one level down. The server compiles, the Inspector calls succeed by hand, and in a live session the agent either ignores the tool or calls it with arguments that do not match the schema.

The failure	What it looks like	What answers it
Built the wrong shape	a skill shelling out to curl, a server wrapping static advice	read both meta-skills first
The skill never triggers	no error, no invocation, no signal	measure the description
One tool per endpoint	technically complete, unusable	tool design as its own step
Passes offline, fails live	Inspector green, session wrong	nothing in this stack — see below

The first fork, which neither item names

Reading both Anthropic skills before writing anything is the actual "should this be a skill or a server" step, and it works precisely because neither one tells you the answer. What they do is make the two costs legible.

Everything in skill-creator is machinery for making a description reliably trigger in natural language. Everything in mcp-builder is machinery for making tool schemas an LLM can call correctly. If your thing is mostly "here is how we do X", the entire apparatus of the second one is overhead you would be paying for nothing. If your thing needs to reach a system that can answer back, the first one has no way to express that. The asymmetry is the answer, and you get it by reading both rather than by finding a page that decides for you.

The gap both of them leave, which nothing here fills

Both items contain a test step, and neither tests the thing that actually breaks.

Skill-creator tests whether the description triggers. Mcp-builder's build phase tests whether the server compiles and answers the protocol. Neither one puts a real request, phrased the way a colleague would phrase it, into a live session and checks which tool got called with what. That gap is where the shipped-and- broken servers live, because it is downstream of every check that is easy to run.

This page used to close that gap with a third item, and it no longer does. What you are left with is the manual version, which is not elegant and does work: after building, open a fresh session with the server attached, ask for the thing in the words a colleague would use rather than the words in your schema, and read the tool call. If it picked the wrong tool, the description is wrong. If it picked the right tool with the wrong arguments, the schema is wrong. Both are fixable and neither is visible from the Inspector.

Two items, and the three that should be here

This stack is short because the honest version of it is short.

There is no row for the MCP Inspector. Mcp-builder tells you to run npx @modelcontextprotocol/inspector, and it is the correct advice. The Inspector cannot be listed here: it is a CLI, and every ingestion path this catalog has produces skills or MCP servers. Same reason Repomix is missing from the codebase-reading stack. The gap is in our shelves, not in the ecosystem.

There is nowhere to publish a skill. Servers have the official MCP Registry and its mcp-publisher CLI. Skills have aggregators, and this project's own sourcing rules exclude them as content sources. That is a real hole in the ecosystem rather than an oversight in this research, and it is worth knowing before you finish writing something you were hoping to distribute.

One item was cut here for the reason the schema stack states: a six-star server that walks you through publishing, which would want registry and CI credentials to do its job. The official CLI does the same work and you build it yourself.

What the search changed

Checked 2026-08-13.

There was a third item on this page: ckorhonen/claude-skills, specifically the mcp-tester skill. It was here to close the live-session gap described above, and the page argued explicitly for keeping a ten-star personal repository in a stack with a hundred-and-sixty-thousand-star one.

Then every stack in this directory went through one grounded question: does any public source describe somebody running this combination, rather than these tools individually. The two Anthropic skills came back paired three separate times:

• named together in anthropics/skills itself;
• described as a sequential MCP-then-skill workflow in a third-party guide that also references skill-creator by name;
• described as a complementary pair, explicitly not as a three-step stack with a tester on the end.

The third one is the finding. A source that names both anchors and stops there is evidence about the shape of the workflow, not just its parts.

For mcp-tester the search found nothing: no post, thread, or repository names it alongside the other two, no per-skill install count, and no commits after 2026-07-05. Official MCP guidance points at mcp-server-dev and the MCP Inspector for the build-and-test half instead.

So the argument in the section above was right about the gap and wrong about having filled it. The gap is real, it is still open, and the honest version of this page says so rather than closing it with an item nobody else has adopted.

npx -y skills add anthropics/skills --skill skill-creator
npx -y skills add anthropics/skills --skill mcp-builder