Founder & CEO, Litchfield Capital Corporation
research-box: an open-source Agent Skill for bundling research work
I just released research-box, an open-source Agent Skill (Apache 2.0) for keeping related research work together: docs, sheets, code, data, chat and agent-task links bundled into a named "box".
What it adds over a folder and a zip:
- Live links stay the source of truth; files are dated snapshots.
- Sharing levels only narrow (private, same assistant, shareable). A drop for wider readers writes a clean copy and says what it left out.
- Every drop carries RO-Crate 1.1 metadata and BagIt-style checksums, so you can tell later whether anything changed.
Before release I tested it on two real research boxes. The very first test caught a privacy leak (box-level notes reaching a shareable drop), which 0.2 fixes. Built with Claude.






Comments4
Paolo, this is genuinely good. I pulled the repo and ran the script myself — the shareable drop actually writes a clean copy, strips the fields that name what you held back, and tells you what got left out. The watch-words check catching a private term inside a file marked shareable is the kind of detail most people skip, and levels-only-narrow is the right call.
Two things worth fixing while it's fresh: an in-place drop without --out still leaves the full box.json and the excluded files sitting in the same folder with just a printed warning — that's the next leak waiting to happen, make --out mandatory for wider audiences. And --check exits 0 even when it finds tampered files, so nothing downstream can actually gate on it.
Solid work though — RO-Crate plus BagIt discipline in a 280-line stdlib script is the right size for what this is. I'll be using it.
Thank you so much for our kind review. I do plan to continue work on it - it is, as you undoubtably have guessed, part of a much larger concept.
Just checked your work - Really interesting stuff!
Thank you for pulling it and running it; that's the best kind of review. You're right on both counts, and 0.2.1 fixes them: the script now refuses any drop that holds something back unless you give --out, and --check exits 1 on any changed, missing or unlisted file (2 with no manifest), so a pipeline can gate on it. Glad you'll be using it. Issues and pull requests are welcome on GitHub.
4 comments
Paolo, this is genuinely good. I pulled the repo and ran the script myself — the shareable drop actually writes a clean copy, strips the fields that name what you held back, and tells you what got left out. The watch-words check catching a private term inside a file marked shareable is the kind of detail most people skip, and levels-only-narrow is the right call.
Two things worth fixing while it's fresh: an in-place drop without --out still leaves the full box.json and the excluded files sitting in the same folder with just a printed warning — that's the next leak waiting to happen, make --out mandatory for wider audiences. And --check exits 0 even when it finds tampered files, so nothing downstream can actually gate on it.
Solid work though — RO-Crate plus BagIt discipline in a 280-line stdlib script is the right size for what this is. I'll be using it.
Thank you so much for our kind review. I do plan to continue work on it - it is, as you undoubtably have guessed, part of a much larger concept.
Just checked your work - Really interesting stuff!
Thank you for pulling it and running it; that's the best kind of review. You're right on both counts, and 0.2.1 fixes them: the script now refuses any drop that holds something back unless you give --out, and --check exits 1 on any changed, missing or unlisted file (2 with no manifest), so a pipeline can gate on it. Glad you'll be using it. Issues and pull requests are welcome on GitHub.