gstack Best Practices: 23 AI Skills and a 7‑Step Workflow for Solo Engineers

What gstack Is

gstack is a collection of AI‑driven engineering tools that turns Claude Code (or any supported AI programming agent) into a complete software‑development team. It is not a framework or plugin; it provides a structured Sprint workflow (Think → Plan → Build → Review → Test → Ship → Reflect) with dedicated AI skills for each phase.

Design Philosophies

Boil the Lake : When AI can achieve full coverage (100% tests, complete functionality), aim for completeness rather than half‑measures.

Search Before Building : Follow a three‑layer knowledge hierarchy (proven patterns, community trends, first‑principles) before writing code.

User Sovereignty : AI offers recommendations, but final decisions remain with the human.

Technical Architecture

Claude Code → CLI (compiled binary) → HTTP POST → Bun Server → CDP → Chromium (headless)

Reasons for choosing Bun over Node.js (from ARCHITECTURE.md):

Compiled binary (~58 MB) eliminates runtime installation.

Native SQLite access reads Chromium’s cookie store directly.

Native TypeScript support removes a separate compile step.

Built‑in HTTP server (Bun.serve()) avoids external frameworks.

The daemon starts a browser in ~3 seconds, then subsequent calls take 100‑200 ms, and it auto‑shuts after 30 minutes of inactivity. Random ports (10 000‑60 000) allow multiple concurrent workspaces.

Ref System

Page elements are referenced with identifiers like @e1, @e2, @c1 using Playwright’s accessibility tree, so the DOM is never modified. This keeps references stable across SPA navigation.

Security Model

Bind only to localhost.

Bearer‑token (UUID v4) authentication.

Cookie handling via macOS Keychain, decrypted only in memory.

Shell‑injection safeguards.

Sprint Workflow Details

Think

Command /office-hours forces the user to answer six product‑direction questions, producing a structured spec in about 15 minutes.

Plan

/autoplan

runs three independent reviews: /plan-ceo-review – CEO perspective. /plan-design-review – Designer perspective. /plan-eng-review – Engineering perspective.

Each review yields an execution plan.

Build

Code is generated by Claude Code (or another agent) and verified with the /browse skill that drives a headless Playwright‑Chromium browser.

Review

/review

simulates a senior engineer, while /codex invokes OpenAI Codex CLI for a second‑model audit, embodying the User Sovereignty principle.

Test

/qa

runs real‑browser tests and can auto‑fix bugs; /benchmark measures Core Web Vitals and page‑load performance.

Ship

/ship

syncs the main branch, runs tests, audits coverage, pushes, and opens a PR. /land-and-deploy merges, CI‑runs, deploys, and validates production health.

Reflect

/retro

records a weekly team retro, and /learn stores lessons for the next sprint.

Practical Modes

Solo Efficient Mode : One developer runs

/office-hours → /autoplan → write code → /review → /qa → /ship

. Reported efficiency gains: scaffold generation ~100× faster, test writing ~50×, feature implementation ~30×, bug‑fix ~20× (based on gstack’s ETHOS.md).

Parallel Sprint Mode : Multiple terminals launch independent Claude Code instances (e.g., /plan-eng-review, /qa, /ship) to run 10‑15 sprints concurrently, as demonstrated by Garry Tan’s 60‑day output of 600 k+ lines.

Cross‑Model Collaboration : Use /codex and /pair-agent to let different large models review each other’s output.

Installation & Team Mode

One‑line install (≈30 seconds):

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup

Team mode adds version synchronization via ./setup --team and gstack-team-init.

Supported AI Agents

gstack works with eight agents (Claude Code, OpenAI Codex CLI, OpenCode, Cursor, Factory Droid, Slate, Kiro, OpenClaw), though full feature parity currently exists only for Claude Code.

Security Mechanisms

Since gstack manipulates cookies and runs code, it includes four‑layer prompt‑injection defenses (content envelope, hidden‑element stripping, data tagging, content filter) and safety commands ( /careful, /freeze, /guard).

Full Sprint Demonstration

/office-hours
/autoplan
# write code with Claude Code
/review
/codex
/qa
/benchmark
/ship
/retro
/learn

This end‑to‑end flow replaces the traditional six‑role pipeline (product, design, front‑end, back‑end, QA, DevOps) with AI‑augmented steps, leaving the human to make final decisions.

Target Audience

Independent developers, small 3‑5‑person teams, open‑source maintainers.

Not ideal for absolute beginners, large enterprises with mature CI/CD pipelines, or non‑web projects.

Conclusion

gstack’s value lies in structuring the entire software‑engineering lifecycle with AI, not in any single feature. The three philosophies guide its design, ensuring AI handles execution while humans retain strategic control. For solo engineers or small teams, a single afternoon trial can replace many manual roles; the MIT‑licensed project can be removed if it doesn’t fit.