Skill types
Every skill is one of three types: Custom prompt, Skill package, or Built-in. You choose the type in the creation wizard, and you can change it later from the skill editor — though packages and built-in skills cannot be converted into each other, so it is worth understanding the differences first.
Custom prompt
A Custom prompt skill is a block of instructions the model follows when the skill is loaded. You write the instructions directly in the Prompt & tools tab of the editor, and Harriet delivers them to the AI whenever the skill is active for that person. If the instructions field is left empty, the skill's description is used instead.
Prompt skills can carry tools too: connect one or more MCP connectors and their tools travel with the skill. That makes this type a good fit for pairing know-how with a system: for example, instructions on how to triage bugs, connected to your issue tracker.
Skill package
A Skill package is a folder of files built around a SKILL.md at the root (YAML front matter with name and description, followed by Markdown instructions) plus optional files such as references/… and small scripts.
Why the format matters: a package is portable. It is the same format Claude uses, so a package you author in Harriet also works in Claude and other Anthropic-compatible tools, and packages published elsewhere can be imported here unchanged. That portability is the whole reason to pick this type over a Custom prompt once a skill outgrows a single page. Files are managed in the Anthropic package tab, uploaded as a zip or a SKILL.md file, or kept in sync from a GitHub repository. The details live on Skill packages.
One rule to know: package files and fallback instructions are mutually exclusive. Either the skill has an explicit SKILL.md file, or it uses its prompt text as the skill content; the editor prevents mixing the two.
Built-in
A Built-in skill wraps a piece of Harriet's own functionality — escalating a conversation into a ticket, searching the knowledge base, looking someone up in your HRIS. You pick one from the Built-in tool dropdown when you create the skill, then add your own instructions on top. The instructions are yours; the tools are not.
Built-in versus a Custom prompt with a connector
Both give the AI tools, so the difference is worth stating plainly: a Built-in skill's tools belong to Harriet, and a connector's tools belong to a system you own.
- Who writes the tools. We do, for Built-in skills. The Built-in tool dropdown is a fixed list that ships with Harriet — you cannot add to it, and you do not maintain what is in it. A connector, by contrast, exposes whatever tools its MCP server offers, and that server is yours (or your vendor's) to run.
- What they can reach. Built-in tools act on Harriet's own data — your tickets, your knowledge base, your synced HR records — which no external connector can reach. Connector tools act on the other system: your issue tracker, your CRM, your file store.
- Who sets up access. A Built-in skill needs no credentials; it already runs as Harriet, subject to the requester's permissions. A connector needs an install, authentication, and someone to keep it working.
- What you change over time. With a Built-in skill you only ever edit your added instructions; we ship changes to the tool itself. With a connector, the tool list changes when the server behind it changes.
A worked example. You want Harriet to raise a ticket when a chat cannot be answered, tag it, and route it to the right inbox. That is a Built-in skill: the escalation tool already knows how to write to Harriet ticketing, and you add instructions about which inbox handles what. Now you want Harriet to open a bug in Jira instead. No built-in tool can do that, because Jira is not Harriet — so you create a Custom prompt skill, attach your Jira connector, and write the instructions around the tools that connector exposes.
When an admin should choose Built-in: the action happens inside Harriet, and a tool for it already exists in the dropdown. If the action happens in another system, or the dropdown has nothing that fits, use a Custom prompt or Skill package with a connector attached.
Comparing the three
| Custom prompt | Skill package | Built-in | |
|---|---|---|---|
| When to use | Playbooks, tone, and process knowledge; quick to author and iterate | Multi-file skills, anything you also want to use in Claude, skills maintained in a repo | Actions inside Harriet itself — tickets, knowledge base, HR records |
| What it contains | Prompt text, plus optional MCP connectors | SKILL.md, optional references/… and scripts, plus optional MCP connectors | One Harriet-maintained tool, plus your own instructions. No connectors. |
| How it updates | Edit the prompt and save | Edit files in the editor, re-import a zip, or scheduled GitHub sync | Harriet maintains the template; you edit only your added instructions |
Not sure which to pick? Start with a Custom prompt. If the skill grows past one page of instructions or needs supporting files, recreate it as a Skill package; the wizard makes that a five-minute job.