nwave-ai/nw-command-optimization-workflow

Command Optimization Workflow

Overview

Convert oversized commands (500-2400 lines) to lean declarative definitions (40-300 lines) by removing duplication, extracting domain knowledge to agents, and adopting the forge.md pattern.

Phase 1: Analyze

1. Classify: dispatcher, orchestrator, or simple
2. Measure: wc -l the file
3. Identify content categories (approximate %): parameter parsing/validation | Workflow/orchestration logic | Agent prompt templates | Examples/documentation | Error handling | Boilerplate (orchestrator briefing, agent registry)
4. Flag reducible content: Duplication from other commands (orchestrator briefing, agent registry, parameter parsing) | Domain knowledge belonging in agents (TDD phases, review criteria, methodology) | Dead code (deprecated formats, aspirational features, old signatures) | Verbose examples (JSON blocks exceeding 3 examples) | Aggressive language (count CRITICAL/MANDATORY/MUST)

Phase 2: Extract

2a: Remove shared boilerplate

These blocks appear in 5-12 commands, extract to shared orchestrator preamble skill:

• "Sub-agents have NO ACCESS to Skill tool" briefing (~20-30 lines/file)
• Agent registry with capabilities (~15-20 lines/file)
• Parameter parsing rules (~10-15 lines/file)
• Pre-invocation validation checklist (~10-15 lines/file)
• "What NOT to include" blocks (~8-12 lines/file)

After extraction, command references preamble.

2b: Move domain knowledge to agents

If command contains HOW the agent works, move to agent definition or skill:

• TDD phase details -> nw-software-crafter | DIVIO templates -> nw-documentarist | Refactoring hierarchies -> nw-software-crafter | BDD/Gherkin -> nw-acceptance-designer | Review criteria -> reviewer agent | Mutation testing config -> nw-software-crafter

Command retains WHAT and success criteria. Agent owns HOW.

2c: Delete dead content

Remove: deprecated format references | Aspirational unimplemented features | Verbose JSON contradicting current format | BUILD:INJECT placeholders | Mixed-language comments

2d: Reduce aggressive language

• "CRITICAL: You MUST extract..." -> "Extract..."
• "MANDATORY: Use the Task tool" -> "Use the Task tool"
• "NON-NEGOTIABLE requirement" -> Remove adjective entirely

2e: Condense examples

Keep 2-3 canonical examples max. Each demonstrates distinct pattern (correct usage, edge case, common mistake).

Phase 3: Restructure

Apply declarative command template (load command-design-patterns skill):

1. Header: wave, agent, overview
2. Context files required
3. Agent invocation with configuration
4. Success criteria
5. Next wave handoff
6. Expected outputs

For orchestrators, add Phases section between overview and agent invocation.

Phase 4: Validate

• [ ] Under size target (dispatchers: 40-150, orchestrators: 100-300)
• [ ] No duplicated boilerplate
• [ ] No domain knowledge belonging in agents
• [ ] No aggressive language (CRITICAL/MANDATORY/MUST count = 0)
• [ ] 2-3 examples max
• [ ] No dead code or deprecated references
• [ ] Declarative structure (WHAT not HOW)
• [ ] Measurable success criteria
• [ ] Clear agent invocation pattern

Phase 5: Measure and Report

Report before/after: Line count (target 60-80% reduction for large files) | Content categories removed/moved | Aggressive language removed | Dependencies created (shared preamble, agent skill additions)

Special Case: develop.md (Mega-Orchestrator)

develop.md at 2,394 lines requires special handling:

1. Embeds 6+ sub-command workflows inline
2. Should reference sub-commands, not embed them
3. Orchestration logic (phase sequencing, resume) is legitimate -- keep it
4. Agent prompt templates for sub-commands -> replace with references
5. Target: 200-300 lines of pure orchestration

Approach: extract orchestration phases as lean sequence (which agent, which command, what context, what gate), remove all embedded sub-command content.

Shared Orchestrator Preamble

Extract common orchestrator knowledge to single shared skill:

• Architectural constraint about sub-agent tool access
• Valid agent list with brief capabilities
• Standard parameter parsing rules
• Pre-invocation validation pattern
• Correct/wrong invocation examples (2 each)

Estimated: ~60-80 lines. Replaces ~620 lines duplicated across commands.