Add prompts and system instructions
You keep pasting the same review instruction into the composer, or the same commit rules, at the start of every session. Put the instruction in a file once and invoke it by name.
Zi has two prompt customizations:
- prompt templates for repeatable slash commands;
- system prompt files for changing the assistant's base instructions.
Start with a prompt template or APPEND_SYSTEM.md. Replace the full system prompt only when you intend to own all base instructions.
Prompt templates
Create Markdown files in:
$HOME/.zi/agent/prompts/
<cwd>/.zi/prompts/Additional files or directories can be listed in the prompts settings array. Project settings paths load first, and only when project configuration is trusted. Global settings paths load before $HOME/.zi/agent/prompts/.
A file named review.md creates /review:
---
description: Review a diff for correctness and risk.
argument-hint: [path]
---
Review $ARGUMENTS.
Focus on correctness, tests, and unintended behavior changes.Arguments preserve Pi-compatible quoting. These forms are available:
$ARGUMENTS- All arguments as one string.
$@- Alias for all arguments.
$1,$2- Positional arguments.
${1:-default}- Positional argument with a default.
${@:2}- Arguments from position 2 onward.
${@:2:3}- Three arguments starting at position 2.
Expansion is single-pass. If no template matches, Zi leaves the slash command text unchanged.
Append to the system prompt
Create:
$HOME/.zi/agent/APPEND_SYSTEM.md
<cwd>/.zi/APPEND_SYSTEM.mdExample:
# Extra rules
- Prefer small patches.
- Ask before adding a dependency.
- If a command fails, show the exact command and failure.Use this when Zi's default prompt is right but you need a few more rules.
Replace the system prompt
Create:
$HOME/.zi/agent/SYSTEM.md
<cwd>/.zi/SYSTEM.mdThis replaces the default prompt. A custom SYSTEM.md owns its product-documentation policy as well as its tool instructions, so use it only when you want complete control.
Precedence
Project resources precede global resources with the same name. That lets a trusted repository set stricter behavior without changing your global setup.
Use AGENTS.md for rules that always matter, a skill for instructions the model should load when relevant, and an extension for executable behavior.