Nightwatch reads an incoming alert and makes the call your on-call engineer would make in the first 90 seconds — then drafts the message. Built for the team without an AIOps budget, where the scarcest resource is uninterrupted sleep.
No narration, no edits. Three alerts in, three different decisions out — the way it would stand your watch tonight.
One to three people share the pager. There's no NOC, no SRE on staff, no AIOps platform — just a phone that buzzes at 3am for a disk-space warning that always clears itself.
So you do the math every on-call does: read the alert, guess the blast radius, decide in 90 seconds whether it's worth waking up for. The decision isn't hard. The volume is what burns you out — and the night it actually matters is the night you've learned to swipe the buzz away.
No "here's what I found, what do you want to do?" An operator decides and routes the work. You come back to a correct page — or a quiet night.
A real fire. Wake a human now, with the blast radius and a first check already drafted.
Real, but it can wait for daylight. Drafted with severity, impact, and a suggested owner.
Known noise or self-healing. Logged with the exact tripwire that would change its mind.
Genuinely your call — rare. It holds a safe default and asks one specific question.
Severity is computed from business impact, not the alert's own label — so a global 0.2% blip can outrank a "CRITICAL," and a "CRITICAL" can be safely muted.
Real outputs — produced by the folder running in Claude, not mockups.
prod-db-1 at 88%. Severity: warning./settings/profile, ~3% of users, stable 20 min, workaround: refresh.org_2290.Put the Nightwatch folder into a Claude Project. Claude becomes the operator — identity, rules, examples, and reference tables all loaded.
Fill the reference/ templates with your services, severity thresholds, SLA customers, and recurring false alarms. ~30 minutes.
It returns one decision, the routed action or drafted message, and a one-line "why" you can audit or override in five seconds.
nightwatch/ ├── identity.md # who it is, what it refuses ├── rules.md # the decision logic (the heart) ├── examples.md # worked calls + edge cases ├── README.md # how to use it └── reference/ ├── severity-matrix.md # blast × impact × trend ├── routing-table.md # sev → who, when ├── known-noise.md # your false alarms ├── sla-customers.md # who overrides the metrics └── response-templates.md# the drafted artifacts
The decision logic is a short-circuiting flow: security and data-loss gates fire first, then noise, then dedup, then severity from the matrix, then the SLA override. The first step that produces an outcome wins.
Nothing is a black box. The rules are plain English in a file you can read, edit, and trust — which is the whole point.
Nightwatch is one worked example. The Operator's Handbook is the method behind it — how to turn Claude into something that decides instead of chats.
Nightwatch is free. There's no account, no agent to install, no dashboard, no per-seat pricing. It's a folder of plain Markdown you drop into Claude — and it's running in a minute. Take it, fork it, make it yours.