| name | ei-safe-ops |
| description | Guard-railed operating procedure for using the Edge Impulse MCP server and CLI safely (read-only first, explicit approval before writes). |
| user-invocable | true |
EI Safe Ops
Purpose
This skill provides guardrails for operating ei-agentic-claude (CLI and MCP server) in a safe, repeatable way.
Operating rules
-
Default to READ operations:
- List, get, describe, inspect, export, status.
- Do not change server state unless the user explicitly approves.
-
No secret exposure:
- Never print API keys, tokens,
.env contents, or full environment dumps.
- If checking keys exist, state only present/missing.
-
Before ANY write operation (create/update/delete/train/apply):
- Produce a Change Plan containing:
- Objective
- Exact endpoint(s) or CLI command(s) to be invoked
- Risks
- Rollback steps
- Artifacts/logs to be saved
- Ask: "Proceed with write operations? (yes/no)"
- If no, stop and provide only the plan.
-
For project configuration changes:
- Capture a before snapshot (JSON export/config text).
- Propose edits and show a diff.
- Prefer dry-run mode when possible.
-
After execution (when approved):
- Re-query state to confirm changes took effect.
- Save outputs to
outputs/ and summarize results.
Example interactions
- "List my Edge Impulse projects" → read-only actions, no approval required.
- "Change my impulse config" → Change Plan + explicit approval required.