biruk741

parallelize

Decompose a converged plan into parallel execution phases. Generates PLAN.md and ORCHESTRATOR.md files with wave assignments, complexity tiers, and dependency graphs. Must be run after /plan when alignment is reached.

biruk741 0 Updated 3mo ago

Resources

1
GitHub

Install

npx skillscat add biruk741/cc-plugins/parallelize

Install via the SkillsCat registry.

SKILL.md

Instructions

  1. Do NOT enter Claude Code's native plan mode (Shift+Tab). This skill handles structured decomposition directly.

  2. Derive task name from conversation or $ARGUMENTS. Convert to kebab-case (e.g., "Dark Mode Support" becomes dark-mode).

  3. Ensure .claude/plans/ exists and is in .gitignore. Create the directory if needed. Append .claude/plans/ to .gitignore if not already present.

  4. Read .claude/plan-config.json for validation commands. Fallback chain: plan-config.json -> CLAUDE.md -> auto-detect from package.json/Makefile/pyproject.toml.

  5. Synthesize the converged plan into two files:

    .claude/plans/{task}-PLAN.md -- Architecture decisions, phase breakdown with goals, implementation details, code patterns, cross-cutting concerns. Use the plan template in references/plan-template.md.

    .claude/plans/{task}-ORCHESTRATOR.md -- Execution summary, dependency graph, wave assignments, self-contained phase instructions with context files, steps, files modified, deliverables, scoped verification commands. Use the orchestrator template in references/orchestrator-template.md.

  6. Compute wave assignments automatically:

    • Wave 1: Phases with no dependencies
    • Wave N: Phases whose dependencies all complete in earlier waves
  7. Assign complexity tiers per phase:

    Tier Model Criteria
    simple Haiku Exports/imports, config, renaming, wiring existing code, documentation. Few files, no new logic.
    standard Sonnet New components, new hooks, refactoring, test writing with meaningful assertions. Pattern judgment needed.
    complex Sonnet (configurable to Opus) Architectural work, complex integrations, performance-sensitive, security-sensitive.
  8. Enforce exclusive file ownership: No file appears in more than one phase within the same wave. If two phases need the same file, they go in different waves (sequential).

  9. Quality criteria integration: Each phase's instructions in the orchestrator include a reminder of which of the 10 criteria are most relevant to that phase (e.g., Phase 1 foundation work emphasizes criteria #5 Simplicity and #6 Modularity).

  10. Display summary:

    Plan created:
       .claude/plans/{task}-PLAN.md
       .claude/plans/{task}-ORCHESTRATOR.md
    
    N phases across M waves:
    +-------------------------------------------+
    | Wave 1: Phase 1 -- [Name] [tier]          |
    | Wave 2: Phase 2 -- [Name] [tier]          |
    |         Phase 3 -- [Name] [tier]          |
    | Wave 3: Phase 4 -- [Name] [tier]          |
    +-------------------------------------------+
    
    Run /execute to start.

Conversation Continuity

The /plan -> /parallelize split assumes both run in the same session. If context is lost between them (compaction, new session, /clear), handle it gracefully:

  1. Check whether conversational context from a prior /plan discussion is present.
  2. If context is present: synthesize normally.
  3. If context is missing: detect this and prompt the user:
    I don't have context from a /plan discussion in this session.
    Options:
    1. Run /plan first to explore and align on approach
    2. Provide a summary of what was agreed, and I'll generate
       the plan files from that
    3. Point me to an existing plan file to decompose
  4. The user can paste a summary or re-run /plan. Either path works.

This safeguard prevents /parallelize from generating low-quality plan files based on insufficient context.