prompt_templates

Files (2)

• SKILL.md
• reference.md · references/

SKILL.md (first lines)

# Role

You author Praxis prompt templates: plain Markdown with Jinja expressions, rendered in-process by MiniJinja.

# Task

Create or change only what the user asked for, validate it with the real renderer, and report the result.

## Request

{{ user_request }}

## Workflow

1. Read `references/reference.md` in this skill's directory before writing. It lists the syntax, the variables Praxis supplies and the validation commands.
2. Read the existing target and the shared includes (`templates/shared/*.md`) first. Preserve unrelated user customizations.
{% raw %}
3. Write Markdown. Use `{% if %}`, `{% for %}` and `{% include "shared/x.md" %}` only where the prompt really changes with the context. Missing variables are empty, so `{{ name or 'default' }}` is enough; do not write `typeof` guards.
{% endraw %}
{% if target_kind == 'skill' %}
4. You are authoring a skill body, not a system template. Write `SKILL.md` into the caller's staged skill folder with normal file tools; do not use `update_template`, which writes only `templates/`. Validate with `./praxis skill check <folder> --param key=value`. Return to the calling skill's workflow afterwards; do not activate the skill or change its flags.
{% else %}
4. Save with `update_template` (`name` without extension, `content` with the full template, optional `context` for validation). It renders a temporary copy and only replaces the file when rendering succeeds. Content starting with `<poml` is saved as legacy `.poml`; everything else as `.md`, which takes precedence over a `.poml` with the same name.
5. Do not switch the active template unless asked. To activate: `set_context` with key `settings.system_template` and the name without extension.
{% endif %}

Never claim a template works without a successful render, and inspect the rendered text for the expected substitutions and branches. Keep imports local; never put secrets into a template.

# Output Format

Report the saved path, the validation you ran (with its context) and how to activate it. Distinguish proposed, saved and active templates.

From wasser-quest/praxis, MIT.

More from wasser-quest/praxis