Import a skill via ZIP

When you already have a ready-made skill — with a manifest, prompt, and scripts — you can bring the whole thing into your organization via Import from ZIP in Skill Studio. This page describes the exact layout the ZIP needs, the limits the validator enforces, and how to import through the interface.

Required layout

The ZIP must contain, at the root, two required files:

my-skill.zip
├── manifest.yaml      # required — Platform schema (schema_version 1)
├── prompt.md          # required — the instructions for the agent
├── requirements.txt   # optional — Python dependencies
└── scripts/           # optional — executable code
    ├── render.py
    └── helpers.py

manifest.yaml (or manifest.yml) — the manifest in the Platform schema. It must have schema_version: 1, a valid skill_key, and a name.
prompt.md — the prompt content (what the agent reads). It's required, even if short.
Any other file (scripts, requirements.txt, assets) is preserved.

Both files are required

A ZIP without manifest.yaml or without prompt.md is rejected. The names are fixed — don't use SKILL.yaml/SKILL.md in the import package.

Limits and validation

Before accepting the package, the validator applies strict rules. Knowing them avoids rejections:

Rule	Limit / behavior
ZIP size	≤ 10 MB. Anything above is rejected.
Number of files	≤ 50 files.
Path traversal	Paths with `..` or absolute paths are rejected.
Symlinks	Not allowed.
Binaries	Extensions `.exe`, `.dll`, `.so`, `.dylib`, `.com`, `.msi` are rejected.
`schema_version`	Must be exactly `1`.
`skill_key`	Must match `^[a-z0-9][a-z0-9-]{1,98}[a-z0-9]$`.
`name`	Required.
`execution_mode`	`local_plain` (default) or `local_protected`. `remote_server` is rejected.
Script scanner	Script files (`.py`, `.sh`, `.js`, `.ts`, `.rb`…) go through a content analysis.

The security scanner

The script files in the package go through a content scan that looks for dangerous patterns — destructive commands, secret-exfiltration attempts, direct calls to model providers, and so on. The result can be:

Blocked — the import fails and the message points to the file and the reason. Fix the script and try again.
Warning — the import proceeds, but Studio surfaces flags for you to review.

The scanner inspects source code, not binaries

Binaries are already barred by the extension rule. The scanner runs over the source code of the allowed scripts. Keep the scripts simple and follow the outputs/ discipline.

Env-secrets consistency

If the manifest declares secrets in both forms — required_env and env.secrets[] — they must agree. A mismatch is rejected on import. See Env-secrets (author) for the contract.

`manifest.yaml` example

A minimal, valid manifest to import, with one declared secret:

schema_version: 1
skill_key: lead-proposal
name: lead_proposal
display_name: "Proposal Generator"
version: "1.0.0"
description: Generates a .docx commercial proposal from a CSV of leads.
entrypoint: scripts/build_proposal.py
permissions:
  network: outbound
  filesystem: outputs_only
docs:
  summary: prompt.md
inputs:
  - name: leads_csv
    type: string
    description: Path to the leads CSV.
    required: true
outputs:
  - name: proposal.docx
    type: file
    description: The generated proposal.
env:
  secrets:
    - name: CRM_API_TOKEN
      secret_ref: crm-api-token
      required: true
execution_mode: local_plain

Step-by-step in the interface

In the web console, open Skills and click Import from ZIP.
Select the .zip file assembled per the layout above.
Choose the destination skill group (required).
Submit. The validator runs the limits and the scanner.
- If something fails, the message indicates what (size, missing file, invalid manifest, blocked script). Fix it and resubmit.
On success, the skill enters as a draft in the editor. Review the Prompt Editor and the Files.
When it's ready, publish — see Publish and version.

Re-importing is safe

Importing the same skill with the same version and the same content is idempotent (a no-op). Changed the content? Bump the version in manifest.yaml, otherwise publishing returns a conflict. See Publish and version.

Was this page helpful?

Report a problem on this pageDo not send passwords, keys, tokens, or customer data.

Required layout​

Limits and validation​

The security scanner​

Env-secrets consistency​

manifest.yaml example​

Step-by-step in the interface​

See also​