Claude Code launches new skill that assembles cross‑functional AI teams on GitHub

Claude Code’s newest skill, dubbed “Assemble,” automates the creation of cross‑functional AI teams directly on GitHub, turning a simple problem statement into a full‑blown project board with dependency‑based execution waves. According to the GitHub repository that hosts the skill, the workflow begins with a virtual Project Manager (PM) that asks four onboarding questions to capture the goal, constraints, scope, and team preferences. Once the intake is complete, the PM pulls from a library of eight predefined teams—Research, Product, Design, Engineering, Infra, QA, Analysis, and Program Management—and assigns each a scoped mission and a set of deliverables, such as markdown files in a /docs folder. The PM then groups the selected teams into sequential waves, ensuring that teams with no interdependencies run first while dependent teams wait for approval of the prior wave’s output (source: GitHub assemble repo).

The execution phase leverages Claude Code’s sub‑agent architecture: for every team in a wave, the PM spawns a dedicated Claude Code sub‑agent that works in parallel, writes its artifacts to disk, and reports status back to the PM. After each wave finishes, the PM pauses for a human checkpoint—“Approve?”—allowing a reviewer to verify the generated research notes, implementation plan, or other deliverables before the next wave proceeds. This pattern of “PM + flat team agents” mirrors Anthropic’s broader vision of AI‑augmented software development, which Wired notes has been reshaping how engineers collaborate (source: Wired).

The skill also embeds a querying interface that lets users ask the PM about any team’s progress, blockers, or the overall project status without launching new agents. For example, a user can type “What is the research team doing?” and receive a concise answer drawn from the board’s state. All artifacts are persisted as markdown files— docs/research‑notes.md, docs/product‑spec.md, docs/implementation‑plan.md, and so on—culminating in an executive summary generated by the PM at project close (source: GitHub assemble repo). This end‑to‑end pipeline reduces the overhead of manual task tracking and enables rapid iteration on AI‑centric projects that would otherwise be bogged down by endless meetings.

Claude Code 2.1.0, which introduced smoother workflows and smarter agents, laid the groundwork for Assemble’s wave‑based parallelism (source: VentureBeat). The new skill builds on that foundation by formalizing the team‑selection process and exposing a plug‑and‑play installation path: developers clone the assemble repository, copy a handful of support files into ~/.claude/skills/assemble, and register the skill via a SKILL.md manifest in the Superpowers plugin directory. Once installed, a single /assemble command launches the PM dialogue, as illustrated in the repository’s example session where a user defines a “CLI tool that analyzes git history” and selects only Research and Engineering teams, skipping Design and Infra. The resulting project board shows Wave 1 (Research) followed by Wave 2 (Engineering), each with task counts and output file paths (source: GitHub assemble repo).

Industry observers see Assemble as a concrete step toward Anthropic’s “cowork” ambition, where AI agents act as collaborative teammates rather than isolated code generators. The Verge reports that Anthropic is pushing developers to adopt Claude for “cowork” scenarios, and Assemble exemplifies that push by handling project management, team coordination, and artifact generation in a single, reproducible flow (source: The Verge). By codifying the often‑implicit choreography of software teams into a deterministic, version‑controlled process, Claude Code aims to cut down the “ad‑hoc staffing and endless meetings” that have long plagued AI projects—exactly the pain point highlighted in the lede.