Claude Code Skill (recommended)¶
The repo ships a companion Agent Skill
for Claude Code —
skill/SKILL.md. It is an
operating guide that loads into the model's context before it drives the MCP, so the LLM acts from
verified facts instead of guesses.
Why it matters: the pt_* tools are powerful, but a model that invents a model name, a port, a
slot id, a cable type, or a raw Script-Engine method will fail — and a bad raw-JS call can even pop
a modal that freezes the live bridge. The Skill encodes the exact, correct facts so that doesn't
happen.
What the Skill covers¶
- A "discover, never invent" discipline — always look up models/ports/modules first.
- The full 43-tool catalog and the mandatory
discover → plan → validate → deployworkflow. - The exact PT Script-Engine API for
pt_send_raw(e.g.ipc.network().getDevice(name)and the globalgetDevices(filter)— there is no globalgetDevice), so raw JS is written correctly. - A wrong → right mistakes table, the 15 cable types, exact port names, the IP/DHCP/routing
conventions, and the module-install slot matrix by router family (HWIC
"0/x", NM"1", NIM"0/1"). - The current known rough edges, so the model works around them.
Install (global — recommended)¶
Run from the cloned repo root. This copies the Skill to your user skills folder so it's available in every project where you use the MCP.
Then run /reload-skills in Claude Code (or restart it). Verify with /skills — you should
see packet-tracer listed. Claude loads it automatically whenever you work with the MCP.
Project-local alternative
Prefer it active only inside one project? Copy the file to that project's
.claude/skills/packet-tracer/SKILL.md instead of the global folder.
Updating¶
The Skill ships with the repo and tracks the MCP version. After git pull, re-run the copy command
above to refresh your installed copy, then /reload-skills.