Claude Code skill adds async shared collaboration sessions

A new community-built skill for Claude Code, called Collab Session, is designed for asynchronous work between multiple human-and-Claude pairs. According to the project description, each participant works in their own Claude Code session and saves progress into a shared workspace without editing anyone else’s files.

The skill is built around a simple rule: every /collab save creates a uniquely named Markdown block file. That means two people can save at the same time without merge conflicts, and the full discussion is assembled later when someone runs /collab join.

The author says the system is meant to preserve the journey, not just the final answer. Instead of storing only a finished decision, Collab Session keeps raw blocks, summaries, and cross-session references so later readers can see what was tried, rejected, and changed over time.

The workflow starts with identity and workspace setup. Users run /collab whoami once per machine, then /collab init once per team to choose a shared transport: either a mounted drive or cloud-synced folder, or a separate mini-repo that uses git as a transport layer.

From there, a participant can start a topic with /collab new, brainstorm in Claude, and save progress with /collab save. The project says secret material can be wrapped in <private>...</private> tags, which are redacted on save so the original content never hits disk.

When a colleague arrives, /collab join pulls the latest state and produces a narrative handoff brief. The description says that brief combines session summaries and recent raw blocks so the next person can pick up without asking which session was active or reloading the entire history.

The project also includes /collab compress, which turns a set of raw blocks into a journey-style summary with contributor attribution, and /collab close, which finalizes a session. After two or more closed sessions, /collab reflect can extract recurring themes into a _patterns.md file.

To keep context windows under control, Collab Session uses tiered retrieval in /collab catchup: a compact index first, then a timeline, then full blocks only if needed. The author says this keeps token cost bounded on long-running work.

Unlike tools built around embeddings or a vector database, the system stores everything as plain Markdown with YAML frontmatter. The project says that makes it compatible with tools like Obsidian, Logseq, Foam, and even grep, while also avoiding model lock-in and rebuild-heavy indexing.

The author says the skill is not just for coding. Suggested use cases include research, market analysis, RFC writing, API design, bug investigation, and refactoring discussions where the reasoning matters as much as the outcome.

The repository also includes hooks that nudge users at session start if they’ve touched an active collab session in the last seven days, and a reminder on session end if they haven’t saved recent work. The project is released under MIT.