Agent skills
Tools tell an assistant what it can do. A skill tells it when — and, more usefully, when not to. Ours are free, open, and written from the things that actually went wrong.
A skill is a Markdown file an AI client reads before deciding how to approach a task. These are published alongside the MCP Session Bridge and work with Claude Code, Claude Desktop, and any other client that reads skills.
signed-in-browser
Working in a browser you are already signed into. The capability is the interesting part, but the guidance is mostly about restraint.
- It says when not to use it. The first row of its decision table is "public page, no login — use an ordinary headless browser". Reaching into a live session for a page you could have fetched is cost with no benefit.
- It leads with the risk. This is a real browser logged into real accounts, so a wrong click sends mail, cancels a subscription or moves money. Read before you write; confirm anything irreversible first; never enter a credential the model generated.
- It says to verify. A click that resolves is not a click that worked. Most silent automation failures are a selector that matched nothing and a caller that never looked.
publish-mcp-server
Listing an MCP server on the public directories — Smithery, Glama, mcpservers.org and awesome-mcp-servers. Every warning in it cost a failed submission to find: a description limit nothing mentions until you submit, a required dropdown that a careless read of the page misses entirely, and paid tiers sitting beside the free submit button.
post-to-hacker-news
Submitting to Hacker News, including Show HN: the title rules, the url-or-text choice, how to tell a real submission from a silent failure, and the author comment a Show HN needs to survive its first hour.
Install one
Drop the folder into your client's skills directory. For Claude Code:
git clone --depth=1 https://github.com/t3ratech/mcp-session-bridge /tmp/t3rnel-skills
cp -r /tmp/t3rnel-skills/agent-skills/signed-in-browser ~/.claude/skills/
Or, for a project rather than your whole account, use .claude/skills/
inside the repository. Skills are plain Markdown — read one before you install it,
which is advice we would give about anybody's.
Why we test them
A skill is an instruction another system follows without checking. One that names a
tool which no longer exists, or a limit that has moved, sends an agent down a path it
was written to rule out. So every session_* tool these name is asserted to
exist, every tool count they quote is compared against what the code actually serves,
and the specific traps are pinned by tests — they fail the build rather than mislead a
reader.