Show HN: Async multi-person collaboration skill for Claude Code

[siclark]

Whilst working with a colleague researching a data model we wanted to build into a website we were struggling sharing the output of our AI sessions, not just the findings but the thought process, the failures as well as success, and any files written.

So I built a Claude skill that enables asynchronous collaboration between multiple human / agent pairs.

Its write once markdown files in a shared location, git, dropbox, network, that all users can view realtime. User driven summaries can be created to save context and summarise status for new joiners to the workspace. If wanted these can be posted to git discussions for wider collaboration or sharing.

On git there are no branches, merges or PRs so nothing to slow you down. Files are uniquely named so no clashes. It's all markdown files with links between summaries and summary commands that extract themes into a patterns.md file. No database, no embeddings or vector DBs. If you want to put Obsidian over the top you can, it's just not built in.

I'd love feedback on the concept and implementation. Raw data kept forever for review, with the compressed summaries available for faster review by humans and to save tokens for agents.

[siclark]

A few extra points I have considered / know about.

Obsidian - Why not just use Obsidian. This works as an Obsidian vault out the box if thats how you want to use it. Summaries use wikilinks so works with the graph view. What this adds is the orchestration and getting data out of the raw conversation in vscode in multiple locations. It captures the journey as well as the outcome.

Why no embeddings / RAG - My pain wasnt searching notes, it was ongoing research and analysis and sharing that live with a friend or colleague without needing to construct a db.

Followups. Single transport per machine constraint. Already in testing and plan to move that to config per workspace.

Happy to chat more if anyone thinks this is useful.

[michelhabib]

2 Claudes better than one :D Did this work better for your team with specific model versions than others?

[siclark]

It did.. the model was teh same, but the prompts the user gave, and so the journey that they took was different to mine. When we brought the knowledge back together it was really interesting, we had a better solution than if I had tried to build out the model and do the analysis myself.

We were researching how to build a model to predict sewage levels in a river that we teach kayaking on.

But also it helped my friend who is less familiar with prompting some good techniques.. he could read my thought process on how I prompted.