superflow

Lightweight Claude Code skill for autonomous product-to-production development. Designed for modern models (Opus 4.6+) — minimal instructions, maximum autonomy.

Philosophy: Many AI coding skills are heavyweight — hundreds of rules that fill context and degrade quality. Superflow trusts the model and keeps instructions minimal. Instead of preventing errors with verbose rules, it catches them through cross-model reviews (Claude + secondary provider). Less context overhead, better output.

How It Works

Phase 1 — You talk, agent proposes. Collaborative product discovery: research, brainstorming, spec, plan. ~15-20 min.

Phase 2 — Agent executes autonomously. PR per sprint, git worktrees, dual-model reviews, TDD. Zero interaction until done.

You: "superflow — upgrade analytics"
Agent: [research → brainstorm → spec → plan] "4 sprints, 28 steps. Go?"
You: "go"
Agent: [Sprint 1 → PR #1] [Sprint 2 → PR #2] [Sprint 3 → PR #3] [Sprint 4 → PR #4]
Agent: "Done. 4 PRs ready, merge in order."

When to Use

Good fit: Multi-file features, new subsystems, architectural refactors (5+ files, needs a plan).

Not a good fit: Quick fixes, config tweaks, single-file changes. Just use Claude Code directly.

Key Behaviors

• PR per sprint — small, reviewable, deployable
• Git worktrees — isolated workspace per sprint
• TDD — write failing test → verify fail → implement → verify pass
• Dual-model reviews — Claude + secondary provider (Codex/Gemini/other); split-focus Claude fallback
• PAR gate — Product Acceptance Review before every push, .par-evidence.json required
• Verification discipline — no claims without pasted test output
• Root cause debugging — investigate before fixing; 3+ failed fixes = question architecture
• Max parallelism — parallelize independent tasks, sequentialize dependent ones

Install

git clone https://github.com/egerev/superflow.git
ln -s $(pwd)/superflow ~/.claude/skills/superflow

Copy superflow-enforcement.md to ~/.claude/rules/ (survives context compaction).

Requirements

• Claude Code CLI
• Secondary provider (optional): Codex (npm i -g @openai/codex), Gemini CLI, or other
• GitHub CLI (gh)
• macOS: brew install coreutils for gtimeout

Files

File	Purpose
SKILL.md	Thin router — startup checklist, phase references
references/phase1-discovery.md	10-step discovery checklist
references/phase2-execution.md	Sprint execution flow
prompts/*.md	Agent templates (implementer, 3 reviewers, testing)
~/.claude/rules/superflow-enforcement.md	Durable rules (survives compaction)

Origin

Originally inspired by Superpowers (Jesse Vincent, MIT). Superflow has evolved into an independent skill optimized for autonomous execution with modern models.

License

MIT