command-authoring

Slash Command Authoring

A slash command is a markdown file at commands/<namespace>/<name>.md. Invoked as /namespace:name [args].

Frontmatter

---
name: <namespace>:<cmd>
description: One-line what the command does. Shows in /help.
argument-hint: "[arg1] [arg2]"   # what $ARGUMENTS will contain
allowed-tools:
  - Read
  - Write
  - Bash
  - Task
  - AskUserQuestion
---

allowed-tools is an allowlist — if set, the command runs with only these. Omit to inherit the session's tools. Prefer allowlists for safety.

Body structure

<objective>
One paragraph: what the command produces and what files it touches.
</objective>

<execution_context>
@${CLAUDE_PLUGIN_ROOT}/<plugin>/references/<doc>.md
@${CLAUDE_PLUGIN_ROOT}/<plugin>/templates/<template>.md
</execution_context>

<context>
Arguments: $ARGUMENTS
Parse as: <describe>
</context>

<process>
Numbered steps. Each step is either:
- Read a file
- AskUserQuestion to fill a gap
- Task-spawn a specialist (by subagent_type)
- Write an output file
</process>

<output>
Files created/modified + a 5-line summary.
</output>

@-imports

@<path> in the body loads the file contents inline at command expansion time. Use ${CLAUDE_PLUGIN_ROOT} for plugin-relative paths. Use this to pull in long reference docs, templates, and few-shot examples so the command body itself stays short.

$ARGUMENTS

Raw string of everything after the command name. Parse it in the body prose — the model handles extraction. For multi-arg commands, document the expected format in argument-hint.

Orchestration commands vs. task commands

• Task command: does one thing inline (e.g., /note, /status).
• Orchestration command: spawns multiple specialists (e.g., /agents-architect:new, GSD's /gsd:plan-phase). Includes gates, approvals, and aggregates specialist output.

Orchestrators should always use the Task tool, not inline the specialist's work.

Design principles

1. Idempotent: running the command twice produces the same result. Check for existing state before creating.
2. Explicit gates: any non-trivial action (writing >1 file, destructive ops) gets AskUserQuestion confirmation.
3. Short body, deep imports: command body under 100 lines; push detail into @-imported references.
4. State on disk: commands that span turns write checkpoint files (.planning/STATE.md, .arch/state.json) rather than relying on context.

References

• @../../agents-architect/references/anthropic-best-practices.md
• @../../agents-architect/templates/command.template.md

Output

Produce commands/<ns>/<name>.md with correct frontmatter, XML-structured body, @-imports for heavy context, and a short exit summary spec.