Local stacks fail in ordinary ways: driver updates, full disks, someone else's batch job, a bad quant, a reverse proxy certificate. Teams that only store tribal knowledge in chat threads lose an hour finding the right person. A one-page runbook does not need drama; it needs order.
What belongs on the page
List the production model tags and the hostnames that serve them. Write the exact commands to check GPU memory, disk free space, and whether the inference process is alive. Include the owner on-call for software versus hardware. Add links to the last known-good flag file and the allowlisted model directory. State clearly what not to do during an incident, such as deleting the entire models folder to "free space" without checking symlinks.
How to keep it true
Review the page on the same cadence you review prompts. After every real incident, add one line about what fooled you. Store the runbook next to your install notes under practices similar to tutorials/install and capacity checks in match/vram-table. Hardware reality still sits at hardware. If you productize access for non-engineers, align the status language with MicroAI One style clarity: say "model loading" instead of a stack trace.
Drills beat optimism
Once a quarter, pick a quiet hour and practice: stop the service, watch the status page, restore from the runbook, and time the steps. If restore needs a second person with a password manager, write that down. For broader delivery and private deployment patterns see solutions and services. Lab machines should not share the production runbook blindly—label lab exceptions with expiry dates.
Extra operational notes
Keep secrets out of the runbook body; point to a vault path. If Mac unified memory hosts are in the mix, link Mac hardware notes so responders do not chase NVIDIA-only commands. Community anecdotes at community are optional reading, not procedure.
Incidents feel slower when the first ten minutes are spent negotiating where the truth lives. Put the truth on one page.
Appendix: what to rehearse in 15 minutes
Walk the on-call through finding the flag file, restarting the proxy, and verifying free disk. Confirm they know which chat channel receives alerts. If Mac hosts are involved, open the Mac hardware notes once during the drill so commands are not assumed to be NVIDIA-only. Add a checkbox for weights volume mounted because missing mounts fake total outages. Keep the appendix short enough to read aloud. After the drill, update the page with one sentence about what confused people. Link capacity checks to the VRAM table so responders do not invent flags under pressure.
Published by MicroAI One (本地AI 1号网), a Shenzhen team working on local LLM hardware, software and deployment.