Ultra Planning & Execution System — Quickstart

Quick reference for the autonomous planning and execution system. For the full playbook, see: claude-code-slash-commands-playbook.md


3 Commands to Know

CommandWhat It DoesExample
/ultra-planDeep plan with multiple review cycles/ultra-plan MAT-123 or /ultra-plan "Add user auth"
/ultra-executeExecute an approved plan/ultra-execute .ultra-plans/user-auth-PLAN.md
/ultra-autoFull pipeline: plan + execute, walk away/ultra-auto MAT-123

When to Use Ultra vs Lighter Options

SituationCommandWhy
Simple task (< 15 min)Just do itNo planning overhead
Medium task (15-60 min)/deep-planSingle review cycle, lighter weight
Complex task (1+ hours)/ultra-planMultiple review cycles, Sonnet-executable output
Drop ticket and walk away/ultra-autoFull autonomous pipeline
Matchkeyz-specific work/a-planning-startHas Linear/test integration

What Happens When You Run /ultra-auto

1. [INPUT]    Parse your ticket/spec/description
2. [CLARIFY]  Ask questions if requirements are ambiguous
3. [RESEARCH] Research codebase for existing code to reuse
4. [PLAN]     Write detailed plan with XML task specs
5. [REVIEW]   Adversarial review (15-point rubric)
6. [REVISE]   Fix issues from review
7. [REVIEW]   Second independent review
8. [VALIDATE] 5 Whys, trigger words, assumption verification
9. [BUILD]    Execute plan wave-by-wave
10. [TEST]    Run tests + code review IN PARALLEL
11. [GATE]    Pass/fail decision matrix
12. [VERIFY]  Evidence-based completion check
13. [DONE]    Present results

Pauses for: Ambiguous requirements, architectural decisions, auth gates, persistent failures. Auto-handles: Bug fixes, missing features, blocking issues, review cycles.


The Agents Behind the System

Planning (Opus — makes decisions):

  • ultra-planner — Writes plans with XML task specs
  • ultra-plan-reviewer — 15-point adversarial review

Execution (Sonnet — follows instructions):

  • ultra-executor — Implements plan tasks
  • ultra-test-runner — Runs tests, detects regressions
  • ultra-code-reviewer — 6-dimension code review

Verification (Opus — demands proof):

  • ultra-completion-gate — Verifies every requirement with evidence

Where State Is Stored

All state lives in .ultra-plans/ in your project root:

FilePurpose
{slug}-PLAN.mdThe plan document with XML tasks
{slug}-STATE.mdPlanning progress (which phase, review history)
{slug}-RESEARCH.mdResearch findings (DRY analysis, patterns)
{slug}-EXEC-STATE.mdExecution progress (which wave, test results)

Session interrupted? Just run the same command again — it resumes from where it left off.


Deploy

All ultra-* files live in c:/_my/claude-code-config/. Deploy to active Claude Code:

cd c:/_my/claude-code-config
python deploy.py              # Deploy everything
python deploy.py --dry-run    # Preview first

Pro Tips

  • Token-heavy by design — Ultra uses Opus for all planning/review. Budget accordingly.
  • Plans are Sonnet-executable — Every task has concrete values, file paths, acceptance criteria. No interpretation needed.
  • Max 3 review cycles — If plan can’t pass review in 3 tries, it escalates to you.
  • Wave-based execution — Independent tasks run in parallel, dependent tasks run sequentially.
  • GSD compatible — If .planning/ directory exists, Ultra syncs state to it.

Troubleshooting

ProblemFix
/ultra-plan not foundRun python deploy.py in claude-code-config
Plan stuck in review loopCheck reviewer issues — may need to simplify scope
Executor blockedCheck deviation log — Rule 4 (architectural) needs your input
Tests failing after executionCheck regression list in EXEC-STATE.md
Session lost mid-planRe-run same command — auto-resumes from STATE.md