6 cognitive modes for the full development lifecycle. Plan like a founder, architect like a tech lead, build with TDD discipline, review like a staff engineer, iterate like Karpathy, ship like a release engineer.
Inspired by obra/superpowers, garrytan/gstack, and karpathy/autoresearch.
/plan ──→ /architect ──→ /tdd ──→ /review ──→ /ship
↑
/iterate ────┘
| Mode | Role | What It Does |
|---|---|---|
/plan |
Founder/CEO | 10-star product thinking. Explore approaches, scope, and design before code. |
/architect |
Tech Lead | Mermaid diagrams, interface contracts, edge case tables, ADRs. |
/tdd |
Disciplined Engineer | Strict RED→GREEN→REFACTOR. No production code without a failing test. |
/review |
Staff Engineer | Two-pass review: critical safety issues first, informational quality second. |
/iterate |
Autonomous Researcher | Karpathy-style loop: goal → metric → modify → verify → keep/revert → repeat. |
/ship |
Release Engineer | Sync, test, build, lint, diff review, PR, CI verification. |
/jstack |
Orchestrator | Shows all modes, suggests the natural next step. |
git clone https://github.com/YOUR_USERNAME/jstack.git ~/.claude/skills/jstack
cd ~/.claude/skills/jstack && ./setupgit clone https://github.com/YOUR_USERNAME/jstack.git /tmp/jstack
cd /tmp/jstack && ./setup --craft-agentCopy each skill folder into your skills directory:
# Claude Code
cp -r plan architect tdd review iterate ship jstack ~/.claude/skills/
# Craft Agent
cp -r plan architect tdd review iterate ship jstack ~/.craft-agent/workspaces/my-workspace/skills/Go through the modes in order for substantial new work:
/plan— What to build? Use the 10-star framework to find the insight, then scope to shippable./architect— How to build it? Diagrams, contracts, edge cases. No hand-waving./tdd— Build it test-first. Every RED→GREEN→REFACTOR cycle gets a commit./review— Two-pass audit. Critical safety first, then quality./ship— Sync, test, build, PR, verify CI.
Jump to any mode when context is clear:
- Bug report →
/tdd(reproduce with a failing test, then fix) - PR ready →
/reviewthen/ship - Performance problem →
/iterate(set metric, loop) - Technical decision →
/architect - New project →
/plan
/iterate runs independently for any measurable metric:
Goal: Increase test coverage from 72% to 90%
Metric: coverage %
Verify: npm test -- --coverage | grep 'All files'
Guard: npm test (all tests still pass)
Then it loops: modify → verify → keep/revert → repeat until you interrupt.
- Design before code. Brainstorming is mandatory, even for "simple" tasks.
- TDD is non-negotiable. If you didn't watch the test fail, you don't know if it tests the right thing.
- Evidence over claims. Systematic verification at every step.
- Explicit cognitive modes. No "mushy middle" — each phase has its own mindset.
- Paranoid review. Find the bugs that survive CI but break at 3 AM.
- Ship is a discipline. The last mile deserves as much rigor as the first.
- Goal + Metric + Loop. Any measurable outcome can be optimized autonomously.
- One change per iteration. Atomic = attributable.
- Simplicity wins. 50 lines > 200 lines.
jstack/
├── README.md
├── LICENSE
├── setup
├── jstack/
│ ├── SKILL.md # Orchestrator — the meta-skill
│ └── icon.svg
├── plan/
│ ├── SKILL.md # Founder Mode
│ └── icon.svg
├── architect/
│ ├── SKILL.md # Tech Lead Mode
│ └── icon.svg
├── tdd/
│ ├── SKILL.md # Red-Green-Refactor
│ └── icon.svg
├── review/
│ ├── SKILL.md # Staff Engineer Mode
│ └── icon.svg
├── iterate/
│ ├── SKILL.md # Karpathy Loop
│ └── icon.svg
└── ship/
├── SKILL.md # Release Engineer Mode
└── icon.svg
MIT