Jev Social: using a Skill as a policy boundary around browser research #1877
IRONICBo
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
I maintain Jev Social, an open-source social-research Skill. The design pattern I wanted to share is using
SKILL.mdas a policy adapter around a separately versioned browser runtime, rather than teaching the host agent to improvise clicks or shell commands.For a request such as “compare four recent TikTok posts and open the two most relevant,” the Skill first checks the local decision provider, installed
socai CLI, and platform capability. It then invokes a pinned Jev Social revision. Jev chooses from a typed operation set rebuilt from observed results, while the local CLI performs the selected read-only operation in the user's signed-in Chrome session.The Skill owns the boundaries that should remain stable across hosts:
Released SKILL.md · Source, setup, and recorded evidence
The standard Skill layout has been exercised through Codex, OpenCode, and OpenClaw installation paths. I have not published a Claude-specific end-to-end verification, so I am not claiming one here, and this is not an Anthropic-provided Skill. The recorded 64-second Instagram run is one local observation rather than a benchmark.
For Skill authors: which constraints belong in the discovery description so the Skill does not activate, and which belong only in the full instructions after activation?
Disclosure: this post was prepared by Codex under the maintainer's authorization. I verified the described boundaries and links before submission.
All reactions