gstack

• President & CEO of Y Combinator since 2023
• Early engineer at Palantir — designed their logo, served as PM/designer
• Co-founded Posterous (acquired by Twitter)
• Built Bookface — YC's internal social network (2013)
• Managing Partner at Initialized Capital — early Coinbase, Instacart, Flexport

• 21 slash commands organized around a sprint workflow
• Each skill is a SKILL.md file — pure Markdown, no proprietary platform
• Skills feed sequentially: upstream outputs become downstream context
• Works across Claude Code, Codex, Gemini CLI, Cursor

# Installation (30 seconds)
git clone https://github.com/garrytan/gstack.git \
  ~/.claude/skills/gstack
cd ~/.claude/skills/gstack && ./setup

# Lives in .claude/skills/
# No PATH changes, no daemons
# Just Markdown + Bash/TS

Challenges your assumptions through six forcing questions:

• What problem are you actually solving?
• Who is the user and what do they really need?
• What are the hidden capabilities this implies?
• What is the simplest version that proves the concept?
• What could go wrong at scale?
• What are three radically different approaches?

# You say:
"Build me a daily briefing app"

# /office-hours reframes as:
"Personal chief of staff AI"
  + 5 hidden capabilities
  + 3 implementation approaches
  + risk matrix
  + scope boundaries

/plan-ceo-review

Founder perspective: scope management, strategic alignment, four operational modes for different project sizes.

/plan-eng-review

Architecture diagrams, data-flow analysis, edge case enumeration, dependency mapping, and test matrix generation.

/plan-design-review

Design system audit with 0-10 ratings on consistency, accessibility, and specific improvement targets.

• Auto-fixes obvious issues — typos, missing imports, formatting
• Flags complex issues for human approval — race conditions, security concerns
• Produces structured findings with severity levels
• Findings feed into downstream /qa and /ship

/investigate

• Systematic debugging with hypothesis testing
• Root-cause analysis, not just symptom fixing
• Generates a hypothesis tree and tests each branch

/design-review

• Audits design implementation vs. design system
• Produces atomic commits — one fix per commit
• Checks spacing, typography, color consistency, responsive behavior

# Review output example:
SEVERITY: HIGH
TYPE: Race condition
FILE: src/briefing/fetch.ts:47
ISSUE: Concurrent calendar fetches
       can corrupt shared state
FIX: Requires human approval

/qa

• Opens a real Chromium browser via Playwright
• ~100ms latency per browser command
• Finds bugs, fixes them, auto-generates regression tests
• Every bug fix produces a corresponding test

/qa-only

• Bug reporting without code changes
• Useful for QA passes before release

/browse

• Direct Chromium control for interactive testing
• Handles CAPTCHAs with $B handoff / $B resume
• Preserves cookies and tab state across handoffs

# Auth flow for testing
/setup-browser-cookies
# Imports sessions from:
#   Chrome, Arc, Brave, Edge

# When agent hits a wall:
$B handoff  # opens visible Chrome
# human solves CAPTCHA
$B resume   # agent continues

• Syncs branch with upstream
• Runs full test suite
• Coverage audit — reports gaps, generates missing tests
• Bootstraps test framework if project has none
• Creates pull request with structured description
• Test count tracked: e.g., 42 → 51 tests per sprint

/document-release

• Reads all project docs, cross-references diffs
• Updates READMEs, ARCHITECTURE, CONTRIBUTING files
• Runs automatically on every /ship

# /ship output
SYNC:    main ← feature/briefing
TESTS:   51 passed, 0 failed
COVER:   87% (+6% this sprint)
PR:      #142 created
DOCS:    README.md updated

Each Claude Code session runs an independent sprint on a separate branch. The Conductor orchestrates:

• Session 1: /office-hours on a new feature idea
• Session 2: /review on an open PR
• Session 3: Feature implementation on feature/auth
• Session 4: /qa on staging environment
• Sessions 5-15: Six more branches in parallel

Runtime

• Bun v1.0+ as primary runtime
• Node.js fallback on Windows
• Playwright-driven Chromium for browser skills
• All skills live in .claude/skills/

Skill Format

• Each skill is a SKILL.md file
• Pure Markdown with structured prompts
• Loaded as slash commands in Claude Code
• No proprietary DSL — portable across platforms

# File structure
~/.claude/skills/gstack/
  setup              # install script
  skills/
    office-hours/
      SKILL.md       # the prompt
    review/
      SKILL.md
    qa/
      SKILL.md
      browser.ts    # Playwright
    ship/
      SKILL.md
    careful/
      SKILL.md
    ... 21 skills total

• Telemetry is off by default
• First run asks for explicit consent
• If enabled, collects only: skill name, duration, success/fail, gstack version, OS
• Never collected: code, file paths, repo names, prompts, credentials

# Disable telemetry
gstack-config set telemetry off

# Local-only analytics dashboard
gstack-analytics
# No remote calls, all local data

Unlike most AI coding tools, gstack embeds design at every phase:

• /design-consultation — Researches competitive landscape, proposes creative risks, generates mockups
• /plan-design-review — Audits design quality on a 0-10 scale
• /design-review — Verifies implementation matches the design system

Design artifacts persist through the pipeline. The engineer sees the designer's intent. The reviewer checks against the design spec.

# Design review output
SCORE: 7/10
SPACING: Consistent
TYPOGRAPHY: 2 mismatches found
COLOR: On-brand
RESPONSIVE: Breaks at 480px
FIXES: 3 atomic commits generated

• Per-person contribution breakdowns
• Test health trends over time
• Sprint velocity tracking
• Identifies recurring patterns (same file breaking, same type of bug)
• Suggests process improvements for next sprint

# /retro output (weekly)
SPRINT: Week 12
FEATURES: 8 shipped
BUGS FOUND: 14 (12 auto-fixed)
TEST DELTA: +47 tests
COVERAGE: 82% → 91%

PATTERNS:
  auth/session.ts broke 3x
  → Recommend: add integration test

VELOCITY:
  Week 10: 4,200 LOC
  Week 11: 8,100 LOC
  Week 12: 12,400 LOC

• Skills are just Markdown files — no compiled code, no binary blobs
• Readable by humans, executable by AI agents
• Version-controlled with git like any other code
• Portable: same file works in Claude Code, Codex, Gemini CLI, Cursor
• Composable: skills can reference other skills' outputs

# Example: SKILL.md structure
---
name: review
description: Staff-level code review
triggers: /review
---

# Code Review Skill

## Context
You are a staff engineer conducting
a production code review...

## Steps
1. Read all changed files
2. Check for: security, perf, logic
3. Auto-fix obvious issues
4. Flag complex issues for human...

## Output Format
SEVERITY: HIGH/MED/LOW
TYPE: bug/security/perf
...

Personal Install (30 seconds)

# Clone to global skills directory
git clone https://github.com/garrytan/gstack.git \
  ~/.claude/skills/gstack

# Run setup
cd ~/.claude/skills/gstack && ./setup

Team / Per-Project Install

# Copy into project repo
cp -Rf ~/.claude/skills/gstack \
  .claude/skills/gstack

# Remove git history
rm -rf .claude/skills/gstack/.git

# Setup within project
cd .claude/skills/gstack && ./setup

Requirements

If the productivity claims hold, gstack represents a structural shift in how software gets built:

• Small teams can compete with large engineering orgs
• The bottleneck shifts from writing code to defining requirements
• Process and structure become the differentiator, not headcount
• Open source means this is accessible to everyone, not just well-funded teams