Notification Templates: Fill Them, Don't Write Them
The video explains the notify-templates skill, which treats notifications as template fills rather than creative writing. It covers the three formats, the routing thresholds, and the rule that sending an alert should take no more than two tool calls.
Every notification Claude sends should cost no more than two tool calls, and none of them should involve writing. That is the point of the notify-templates skill, covered in this Humanitarians AI video narrated by Liam. The skill treats notifications as template fills rather than creative writing, and it explicitly says not to spawn a subagent for the job. The video opens by correcting a common assumption: that Claude writes each low-stock alert from scratch.
What a skill is and what this one contains
The video describes a skill as a folder Claude reads before it acts. This one holds three fixed formats: a low-stock Slack alert, a supplier email, and an escalation for human review. Claude fills the slots from data it already has, appends the result, and stops. Because the data, the template and the append are the same every time, there is no creative step to delegate.
Routing instead of a fixed destination
Where a message lands is also decided by rules rather than by Claude's judgment in the moment. The video lists the tiers:
- Most notifications go to the ops channel by default.
- An active stockout on a top SKU adds an @here.
- Anything over $25,000 goes straight to a purchasing lead instead of the channel.
- Finance only hears about it past $100,000 outstanding, or when a duplicate order is suspected.
Cross a threshold, and the message says which one was crossed.
Batch, do not spam
For a daily sweep across many SKUs, the rule is one summary message, not one per SKU. This holds even when the task asks for a note to each SKU: every line still goes into a single batch append. Sending is one append to an outbox file in JSON, with one line per notification.
The video gives a simple test for whether the process has gone wrong. If sending an alert takes more than two tool calls, something has been over-engineered.
The carry-out
The video's summary is short. Claude does not write these alerts; it fills three fixed templates from data it already has. A sweep becomes one summary and one append, never one call per SKU.
Key takeaways
- Notifications are template fills, not creative writing, and the skill says not to spawn a subagent for them.
- The three fixed formats are a low-stock Slack alert, a supplier email and an escalation for human review.
- Most alerts default to the ops channel, and an active stockout on a top SKU adds an @here.
- Amounts over $25,000 go directly to a purchasing lead rather than the channel.
- Finance is contacted only past $100,000 outstanding or for a suspected duplicate order.
- A sweep across many SKUs produces one summary and one batch append, even if a note to each SKU was requested.
- Sending a notification should never take more than two tool calls.
Try it yourself
The video gives a prompt to paste into Claude. It asks for a low-stock alert for SKU-0042 with 3 units on hand, a reorder point of 20 and 2 days of cover, and notes that eight other SKUs also dropped below their reorder point that day. It asks Claude to read the notify-templates skill and explain what it will send, one alert or one for each of the nine, and where each is routed, before sending anything. Run it against your own notification scenario.
Chapters
Full transcript(auto-generated, with timestamps)
The naive framing: "does Claude write the alert template?"
[0:00]Someone assumed Claude writes each low stock alert from scratch. It doesn't, it fills a fixed template. So, does Claude fill the low stock alert template? A
Anatomy: three fixed templates, no subagent
[0:09]Skill is a folder Claude reads before it acts. This one is notify templates, three fixed formats, a low stock slack alert, a supplier email, and an escalation for human review. Fill the slots from data already at hand, append the result, and stop. The file says explicitly, do not spawn a sub agent for this.
Mechanism: where the message is routed (ops / purchasing / finance)
[0:26]Where the message lands isn't fixed either, it's routed. Most things go to the ops channel by default. An active stockout on a top SKU adds an at here. Anything over $25,000 goes straight to a purchasing lead, not the channel. Finance only hears about it past $100,000 outstanding or a suspected duplicate order. Cross a threshold and the message says which one. For a daily
Batch, don't spam — one summary, one append, ≤ 2 calls
[0:47]Sweep across many SKUs, the rule is one summary message, not one per SKU, even when the task asks for note to each. Every line still goes into a single batch append. Sending it is one append to an outbox file in JSON, one line per notification. If sending an alert takes more than two tool calls, something's gone wrong. Same data, same template, same append every time. Claude doesn't
Carry-out
[1:08]Write these alerts, it fills three fixed templates from data it already has. A sweep becomes one summary and one append, never one call per SKU. Your turn, paste this into Claude. I need a
Your turn
[1:19]Low stock alert for SKU 0042, three units on hand, reorder point 20 days of cover two. Also, eight other SKUs dropped below their reorder point today. Read the notify template skill and walk me through what you'll send, one alert or one for each of the nine, and where each one is routed before you send anything. Fill the template, don't write
Outro
[1:38]It. Lay 'em in for bear.
More from Claude
2:00Claude's Legal Hold Skill: One File, Four Modes, Zero Judgment
1:25What's Actually in a Claude Skill? A One-Sentence Answer
1:50Claude's "No Rewrite, Ever" Rule: Inside the Legal Writing Skill
2:45The 402-Line Prompt: How Splitting Agents Makes Them 5x Faster
2:03Claude Picked Up a Skill Called "log-leave" — Can It Approve Leave Requests?
1:58