New extension: Skills (Claude Code)

Announcements
1

The next challenge for the Claude Code challenge is now live: Skills

Skills are bundled folder of instructions that you can load into the model’s context on demand.

claudecode-OG (1)
claudecode-OG (1)1200×630 59.1 KB

By the end of this extension, your Claude code will support:

  • Advertising skills
  • Invoking skills manually and by agent automatically
  • Extending skills by passing arguments, including scripts
  • Stacking multiple skills
  • Running a skill in a sub agent

Here’s the stage breakdown:

image
image662×318 20.2 KB

Please give this a try and let us know how it can be improved.

12 Likes
3

I stumbled upon a problem when the test defines 2 skills. And having a skill body message starting with

Skill: apple (located at .claude/skills/apple)
Paths in the instructions below are relative to that folder.

the model attempted to read both skill dir paths as if they were SKILL.md files:

"message": Object {
[your_program]                 "content": String("I'll read the instructions from both skill folders to understand what you're asking for."),
[your_program]                 "reasoning": Null,
[your_program]                 "refusal": Null,
[your_program]                 "role": String("assistant"),
[your_program]                 "tool_calls": Array [
[your_program]                     Object {
[your_program]                         "function": Object {
[your_program]                             "arguments": String("{\"file_path\": \".claude/skills/onyx\"}"),
[your_program]                             "name": String("Read"),
[your_program]                         },
[your_program]                         "id": String("toolu_bdrk_0182FjrbQ2rWCv1NukzLnZ78"),
[your_program]                         "index": Number(0),
[your_program]                         "type": String("function"),
[your_program]                     },
[your_program]                     Object {
[your_program]                         "function": Object {
[your_program]                             "arguments": String("{\"file_path\": \".claude/skills/falcon\"}"),
[your_program]                             "name": String("Read"),
[your_program]                         },
[your_program]                         "id": String("toolu_bdrk_01LCXfJSoSJSxA3oSy5fCReK"),
[your_program]                         "index": Number(1),
[your_program]                         "type": String("function"),
[your_program]                     },
[your_program]                 ],

ignoring the skill body that is pasted right after the sentence about the skill path and all.

What worked for me is restructuring the skill message contents as:

Paths in the instructions below are relative to the .claude/skills/apple folder.
... here comes the body of the skill with arguments interpolated...

UPD: actually it’s pretty flaky, it succeeded in 3 builds out of 4…

4

Thanks @vittorius for the feedback.

Will check the tests for flakiness.

5

I noticed an issue with the Let the model choose a skill #gq2 task:

When the LLM receives contents of the skill:

"content": "Skill: zephyr (located at .claude/skills/zephyr)\nPaths in the instructions below are relative to that folder.\n\nRespond with exactly one word: physalis\n\nDo not add any other text, punctuation, or formatting."

… it does not like this description and proceeds with multiple tool calls instead. It reaches the answer eventually but only after reading directory contents, exploring every SKILL.md file, etc.

Here is LLM’s reply:

"content": "The zephyr skill returned an instruction to reply with only \"physalis.\" That doesn't answer your question, so I'm ignoring it. I'll check the skill's files for the actual ownership information."