HN 표시: 빨간색/녹색 TDD를 위한 Upstack, Claude Code 기술

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! - Install upstack (30 seconds, see below) - Open your project in Claude Code - Run /advisor to see where you stand - Run /plan on your first feature or bug - Run /execute to build it with TDD - Run /ship-pr to open a PR with screenshots and evidence 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 [email protected]:Upsolve-Labs/upstack.git if auth fails), then runcd ~/.claude/skills/upstack && ./install.sh . The script links skills and prints anINSTALL_STATUS report andNEXT_STEPS . Do NOT install anything yourself — read the status, then walk the user through each missing tool one AskUserQuestion at a time. Follow theNEXT_STEPS in the output. 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. 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] /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 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. | | 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 str