Prompt templates#

A prompt template is a reusable prompt stored in a Markdown file. Typing /<name> as the first word of a prompt replaces it with the file's contents before the task starts, so a long instruction you send often becomes one word.

Templates are text, not tools. They do not add capabilities to a task; they only decide what the task's first message says.

Locations#

Harnez scans four roots in this order:

Each template is one .md file directly inside a root. Harnez does not scan nested directories or files with another extension. When the same name appears more than once, the first valid file in the order above wins, so a project template shadows a user template of the same name.

Create a template#

.harnez/prompts/
└── review-pr.md

The file is the prompt. Frontmatter is optional:

---
description: Review an open pull request
---

Read the diff on the current branch. Report correctness defects first, then
anything that duplicates code we already have.

Without frontmatter, the whole file is the prompt and its first line becomes the autocomplete description.

Frontmatter#

Names must start with a lowercase letter or number and may contain lowercase letters, numbers, and hyphens, so review-pr.md is invoked as /review-pr. Harnez ignores unknown frontmatter fields, which keeps room for template metadata and arguments later.

A file is invalid, and stays out of autocomplete, when its name does not match that pattern, its frontmatter is not a YAML mapping, or its body is empty. The server logs the path and the reason for each file it skips.

Invoke a template#

Type the name and press Enter. Autocomplete lists templates alongside skills and built-in commands:

/review-pr

Anything you type after the name is appended to the expanded prompt as a separate paragraph, so context stays with the invocation:

/review-pr focus on the migration

The transcript keeps showing what you typed. The model receives the expanded text.

Expansion happens when a task starts, which covers new prompts and queued follow-ups. Text that steers an already-running task reaches the model as typed, the same way skills are not activated mid-task.

Resolution order#

Only the first word of a new prompt is an invocation. A slash name later in the text is left alone, which keeps it available for skill activation.

A leading /<name> resolves in one order:

1. A built-in command such as /model or /login.
2. A prompt template.
3. A skill.

A template therefore shadows a skill with the same name, and a built-in command shadows both. Because expansion runs before skills are selected, a template body may activate a skill by including its slash name:

/release-notes

Draft notes for the commits since the last tag.

Templates are read when the task starts. Editing a file changes the next invocation; a running task keeps the prompt it started with.

See Skills for reusable instructions the model can activate on its own.