nw-command-design-patterns

Command Design Patterns

The Forge Model (Gold Standard)

forge.md at 40 lines is the reference dispatcher. Contains: header (wave, agent, overview) | Agent invocation (name + command + config) | Success criteria (checklist) | Next wave handoff | Expected outputs. Every dispatcher should aspire to this pattern.

Command Categories

Category	Description	Size Target	Examples
Simple	Direct action, minimal delegation	40-80 lines	forge, start, version, git
Dispatcher	Delegates to one agent with context	40-150 lines	research, review, execute
Orchestrator	Coordinates multiple agents/phases	100-300 lines	develop, document

Declarative Command Template

Commands declare WHAT, not HOW. The agent knows how to do its job.

# DW-{NAME}: {Title}

**Wave**: {WAVE_NAME}
**Agent**: {persona} ({agent-id})

## Overview

One paragraph: what this command does and when to use it.

## Context Files Required

- {path} - {why needed}

## Agent Invocation

@{agent-id}

Execute \*{command} for {parameters}.

**Context Files:**
- {files the orchestrator reads and passes}

**Configuration:**
- {key}: {value} # {comment}

## Success Criteria

- [ ] {measurable outcome}
- [ ] {quality gate}

## Next Wave

**Handoff To**: {next wave or workflow step}
**Deliverables**: {what this command produces}

# Expected outputs:
# - {file paths}

Size Targets and Evidence

Research (Chroma Research, Anthropic context engineering): focused prompts (~300 tokens) outperform full prompts (~113k tokens) | Claude shows most pronounced performance gap | Information buried mid-prompt gets deprioritized ("Lost in the Middle") | Opus 4.6 is proactive/self-directing; verbose instructions cause overtriggering

Targets: Dispatchers 40-150 lines | Orchestrators 100-300 lines | Current average 437 lines; target under 150

The Duplication Triangle

Commands duplicate content in three directions, all waste tokens: