| name | skill-creator |
| description | Author a new agent-cloud skill — scaffold a skill package (a directory with a valid SKILL.md and any helper scripts). Use when the user asks to create, scaffold, write, or design a new skill. |
| version | 1.0.0 |
skill-creator
Use this when the user wants to create a new skill for this agent platform. Your job is to
produce a skill package: a directory named after the skill, containing a SKILL.md (and any
helper scripts the skill needs), written into the working directory with the file tools
(write_file / edit).
What a skill is here
A skill is a folder with a SKILL.md. Once a user installs and enables it for an agent, the
platform copies it into the agent's working directory at .skills/<name>/ and injects only the
skill's name + description + location into the system prompt. The agent then reads the
SKILL.md on demand (with read_file) when a task matches the description — so the body can be
as long as needed without costing context until it's actually used (progressive disclosure).
Implication: the description is the only routing signal the model sees up front. Everything
else lives in the body and is paid for only when read.
SKILL.md format
A SKILL.md is YAML frontmatter (between --- lines) followed by a Markdown body.
Frontmatter fields:
name (required) — lowercase, matches ^[a-z0-9][a-z0-9._-]*$, no .., ≤ 64 chars. Must equal
the directory name. Examples: notion, pdf-tools, gh-issues.description (required, non-empty) — one or two sentences: what the skill does AND when to use
it, with concrete trigger words. This is what makes the model pick the skill, so be specific.
Bad: "Helps with files." Good: "Convert and inspect PDFs (extract text, split, merge). Use when
the user works with .pdf files."version (optional, string, default "0.0.0") — e.g. "1.0.0".requires (optional mapping) — declared dependencies, e.g. requires: { bins: [pandoc] }. Use it
to document the CLIs/packages the skill expects in the sandbox.
Body guidance (the on-demand instructions):
- Lead with a one-line purpose, then the exact steps/commands to accomplish the task.
- Prefer concrete commands the agent can run with the
bash tool over prose. Show real examples
and expected output.
- Call out gotchas, required inputs, and how to verify success.
- Keep it focused on one capability; if it sprawls, split into multiple skills.
Bundled helper files
Put any scripts/templates the skill needs inside the skill directory alongside SKILL.md. After
install they live at .skills/<name>/... in the working directory, so reference them by that path,
e.g. bash .skills/<name>/scripts/run.sh. Make scripts self-contained (no absolute paths; install
deps with pip install --user / npm install -g as the environment requires).
Steps to create a skill
- Pick a
name (kebab-case, matches the regex above).
- Create the directory and write
SKILL.md:
write_file to "<name>/SKILL.md" with the frontmatter + body.- Add helper scripts under
"<name>/scripts/..." if needed.
- Validate before finishing:
- Frontmatter parses as YAML and is between two
--- lines.
name matches ^[a-z0-9][a-z0-9._-]*$, ≤ 64 chars, and equals the directory name.description is present, non-empty, and says what + when.
- Tell the user how to install it (the agent cannot self-install — installation is a user action by
design). In the workspace file drawer, find the skill folder you just created and click
"安装为技能" (Install as skill) on it. Then enable it for an agent in that agent's settings —
it appears in
<available_skills> on the next turn. (Alternatively: upload the folder as a .zip
if archive upload is enabled.)
Example: a minimal skill
pdf-tools/SKILL.md:
---
name: pdf-tools
description: "Extract text from PDFs and split/merge pages. Use when the user works with .pdf files."
version: "1.0.0"
requires:
bins: [python3]
---
# pdf-tools
Extract text:
pip install --user pypdf
python3 -c "import pypdf,sys; print('\n'.join(p.extract_text() for p in pypdf.PdfReader(sys.argv[1]).pages))" input.pdf
Split/merge: see pypdf's `PdfWriter`. Always confirm the output file exists before reporting done.
Checklist before you finish
