upstack

Lightweight Claude Code (+ others) skill suite, used by the team at Upsolve AI.

Red/Green TDD workflow with 9 focused skills, minimal dependencies. Inspired by gstack and simon wilson.

Hi, I'm Serguei, CTO at Upsolve. I built Hyperauto at Palantir and I'm now working to help teams build robust, reliable data agents for their teams and customers. Like Garry, I think we're currently on the cusp of something huge in the AI coding space. But after playing around with gstack, I found its approach to be much more suited for new & ambitious greenfield projects than iterating on existing products.

At Upsolve, we are hyper focused on security, reliability, and grounding agents in real, auditable data. So our team's engineering approach requires a slightly different approach. With code being essentially free to write now, I believe we developers now need to focus on providing tightly scoped specs, project definitions, and test suites before touching a line of code. When you use upstack, it forces you to think test-first and submits code that is instantly provable and reliable. It takes 30 seconds to set up, so give it a try!

Quick Start: Your First 5 Minutes

1. Install upstack (30 seconds, see below)
2. Open your project in Claude Code
3. Run /advisor to see where you stand
4. Run /plan on your first feature or bug
5. Run /execute to build it with TDD
6. Run /ship-pr to open a PR with screenshots and evidence

Install — Takes 30 Seconds

Install on your machine

Open Claude Code and paste this. Claude does the rest.

Install upstack: clone https://github.com/Upsolve-Labs/upstack.git to ~/.claude/skills/upstack (try HTTPS first, fall back to SSH with git@github.com:Upsolve-Labs/upstack.git if auth fails), then run cd ~/.claude/skills/upstack && ./install.sh. The script links skills and prints an INSTALL_STATUS report and NEXT_STEPS. Do NOT install anything yourself — read the status, then walk the user through each missing tool one AskUserQuestion at a time. Follow the NEXT_STEPS in the output.

Codex, Gemini CLI, or Cursor

upstack uses the SKILL.md standard. Clone the repo, then copy the skills into your tool's skill directory:

git clone https://github.com/Upsolve-Labs/upstack.git ~/.claude/skills/upstack

# Codex
cp -r ~/.claude/skills/upstack/skills/* .agents/skills/

# Cursor
cp -r ~/.claude/skills/upstack/skills/* .cursor/skills/

# Gemini CLI
cp -r ~/.claude/skills/upstack/skills/* .gemini/skills/

Skills are plain markdown — they work in any agent that reads SKILL.md files.

See It Work

you:    I want to build a telemetry dashboard for my SaaS app
you:    /plan
claude: [audits codebase, compares 2-3 alternatives, maps error paths, test coverage diagram, implementation plan, structured tickets with dependency DAG, offers to materialize to Linear/TODOS.md]

you:    /execute
claude: [navigates your app with agent-browser, writes failing tests, implements, confirms green]
        or: /execute --ticket P1-3
claude: [reads ticket brief, checks dependencies, scopes tests to acceptance criteria, implements, marks done in TODOS.md]

you:    /validate
claude: [walks through every planned path manually, screenshots UI, saves API examples to evidence/]

you:    /review
claude: [diffs against base branch, finds bugs and security issues, numbered findings like 1A, 2A]

you:    /ship-pr
claude: [writes docs, bumps version, opens PR with screenshots and Linear links]

Workflow

/plan -> /execute -> /validate -> /review -> /ship-pr

Skill	What It Does
/plan	Purpose, scope, eng review, test proposal, implementation proposal, structured tickets with dependency DAG, optional Linear materialize. Tests before code.
/execute	RED: write failing tests. GREEN: implement. Atomic commits. Supports --ticket <ID> for ticket-scoped execution with dependency checks and scope guardrails.
/validate	Manually verify every path works. Save screenshots + API examples to evidence/.
/review	Senior engineer code review against base branch. Best in a clean conversation.
/ship-pr	Docs, version bump, PR with evidence screenshots and Linear links.

Automated flow:

Skill	What It Does
/upstack-run	Full sprint: plan -> execute -> validate -> review -> ship-pr. Loops until clean, ends with a pushed PR.

Independent skills:

Skill	What It Does
/qa-review	Exploratory testing: traverse UI with agent-browser, find edge cases, screenshot findings.
/advisor	Analyze git state and recommend which skill to run next.
/setup	Check prerequisites and verify installation.
/upgrade	Pull latest upstack with backup.

Dependencies

Tool	Required?	Why	Install
gh	Yes	/ship-pr uses it to push commits, create/update PRs, and generate release notes.	macOS: brew install gh · Linux: see docs · Windows: winget install --id GitHub.cli
agent-browser (by Vercel)	Yes	/plan, /validate, /review, /qa-review use it to navigate frontend, click around the browser, and screenshot functionality.	macOS: brew install agent-browser · Other: npm install -g agent-browser
linear-cli	Optional	/plan materialize phase creates Linear tickets from structured plan output. /execute updates ticket status during work. Falls back to TODOS.md when unavailable.	macOS: brew install schpet/tap/linear-cli · Other: npm install -g linear-cli

The install script detects your OS and what's missing, then walks you through each one.

Upgrade

/upgrade

Or manually: cd ~/.claude/skills/upstack && git pull

Philosophy

• Tests before code, always. Red/Green TDD is the default path.
• Evidence-based PRs. Screenshots and API examples ship with every PR.
• Scope discipline. Out-of-scope discoveries go to TODOS.md (or Linear tickets), not into the current work.
• No bloat. 10 skills, ~450 lines total. No build system, no templates, no binaries.

License

MIT License. Free and open source.