Skilify

Skilify is a built-in skill that teaches the AI on your desktop to turn something you've done successfully into a reusable skill package — and to submit it to your organization's review queue, so teammates can adopt it after approval.

What Skilify does

When someone completes something useful in Harriet Desktop (a repeatable checklist, a research pattern, a way to drive a connector), Skilify guides the assistant through capturing it as an Anthropic-compatible package: a folder built around SKILL.md (YAML front matter with name and description, plus Markdown instructions), with optional references/… files beside it. The format is the same one described on Skill packages.

Skilify insists on testing before sharing: run the skill end-to-end on a clean session using only what is in the folder, fix the gaps, and only then submit. It also draws a hard line on content: skills are instructions and conventions, not secrets, so tokens, API keys, and private customer data never belong in the package.

Skilify is created automatically for organizations with Endpoint AI. Its channel defaults are off and it is not added to the default profile, so an admin decides who gets it; see How skills reach people.

How submissions work

Skilify submits into the review queue, never straight to your live catalog: there is no auto-publish path, even for account owners. Before anything is uploaded, the assistant must have the author's explicit agreement: the package will be visible to reviewers and stored for moderation and possible publication to teammates.

There are two credential paths:

• From an enrolled device: the assistant uses the device credentials Harriet already provisioned (device_id, api_key, and server in ~/.harriet/config.json), typically through the list_skill_packages and submit_skill_package tools.
• With an upload key: no device enrollment required. Generate a short-lived key from My AI; it is an hss_… token shown once, and you can revoke keys from the same screen. The assistant submits over HTTPS with that token.

Either way, the assistant runs a duplicate check first; Harriet may reject a submission whose skill name already exists for your organization. If neither path is available, the fallback is manual: create a package skill in the console, paste the files, and use the normal Save and submit for review path.

A power user's typical flow

1. Build the workflow locally. Iterate in Harriet Desktop until the outcome is reliable; the successful session is the raw material.
2. Ask the assistant to skilify it. With Skilify assigned, the assistant drafts SKILL.md: the outcome, the numbered procedure, inputs the user must provide, tools to call, and guardrails.
3. Test on a clean session. Confirm the packaged instructions reproduce the success without hidden context from the original conversation.
4. Consent and submit. Agree to share, and the assistant runs the duplicate check and submits. The package lands in the review queue with status Submitted.
5. Review and rollout. A reviewer approves (see Review workflow), then an admin assigns the skill to profiles, teams, or people.