A simple connector between your Nightscout data and the AI assistant you already use
Say “Check my levelz” and it reads your numbers first, then answers like an empathetic diabetes educator who still gets to the point.
Steady afternoon so far, nice.
Sample reply · numbers are examples
Levelz doesn't just go on what you tell it. It knows the full context of your day and your data: your current reading and where it's trending, how today has gone so far, the insulin and carbs already working, and your loop's forecast.
Levelz is a set of instructions you add to your AI assistant. It does two jobs. In the moment, whenever you mention your sugar, a low, carbs or what you're about to eat, it pulls your live Nightscout data first: your number, trend, insulin and carbs on board, your loop's forecast and your carb ratio. Over time, ask it to review your settings and it looks for patterns in the last two weeks and suggests which basal rates, carb ratios or correction factors may be worth adjusting with your care team.
Every reply opens with a short, honest read on how your day is going. Anything marked is a prediction from your loop's forecast, and marks advice. It picks its format for the moment: a quick glance and numbered steps for everyday checks, two lines when you're low, a chart and change table for reviews, and plain spoken sentences when you're using voice.
These are just examples. There are no magic words. Claude understands plain language, so you can say “I'm at Miss Likklemore's, what should I get?” and it will look up the restaurant and its menu, then check your numbers before suggesting what to order.
Sample numbers, so you can see how Levelz answers in different moments.
Steady afternoon so far, nice.
1.6 U on board · 0 g carbs · ratio 1:12 tonight
More insulin is working than food, so start dinner with some carbs rather than waiting on the mains.
About 45 g is roughly 4 units at 1:12. See if your loop agrees.
🟢 in range · 🟡 heads-up · 🔴 act now · prediction · advice
🔴 2.8 and falling. Let's get you back up. Have 15–20 g of fast sugar now: juice, glucose tabs or regular pop.
Where it's heading: 1.2 U of insulin is still working against 0 g of carbs on board, and Loop has paused your basal. Without sugar, Loop sees you near 2.5 in 30 minutes.
Next: recheck in 15 minutes and repeat if you're still under 4. Don't drive until you're above 4 and holding.
Once you're up: that walk earlier is probably still working, so have a small snack before dinner.
“Steady afternoon so far. You're at 6.4 and steady. Looking ahead, with more insulin working than food, you'll likely drift low within the hour. Start dinner with some carbs. Next, have a small snack now, order the jerk chicken with half a side of rice, and check again 30 minutes after you eat.”
Long afternoon, I know.
2.1 U on board · 0 g carbs · pod site 3 days old
You already have a good amount of insulin working, so another correction risks stacking. A brisk 15-minute walk is a good bet if ketones are clear.
Time in range 78% · lows 6.2% · highs 15.8% · estimated A1c 6.8%
Two solid weeks overall, with one pattern worth fixing.
Fix lows first: they cluster late morning and around lunch, and about half your lunches end in a low a few hours later. Your loop is also quietly cutting your afternoon basal by about 40%, which usually means it's set higher than you need. Leave the evening highs for the next round.
Changes to discuss with your care team| Setting | When | Now | Try | Why |
|---|---|---|---|---|
| Carb ratio | 11:00–16:00 | 1:10 | 1:11 | 48% of lunches followed by a low 2–5 h later |
| Basal | 12:00–18:00 | 0.80 U/hr | 0.70 U/hr | Loop delivered only about 60% of scheduled basal |
Here's what I use. Only two pieces are essential: Nightscout and an AI assistant. Everything else can be swapped for whatever you already use.
CGM → loop → Nightscout → Levelz in your AI assistant
Your CGM or loop data already flowing into it. Self-hosted or a hosted service both work. I use Nightscout Pro: it's fully managed, so there's nothing to host or update yourself.
In Nightscout, Admin Tools → add a subject with the readable role, then copy its token. On a hosted service, if you can't find Admin Tools, ask their support for a read-only token.
Claude: a paid plan (Pro or higher). The Free plan can't allow Claude to reach your Nightscout address. ChatGPT: a paid plan, since making your own GPT needs one. The account holder should be an adult; for a child, a parent sets it up.
Two steps, about ten minutes.
Claude's workspace is locked down by default, so allow just your address.
Download your personalized skill and upload it: in Claude, go to Customize → Skills → + → Create skill → Upload a skill, and choose the zip. Use the zip if you can: the scripts come with it, so every check runs straight away.
No way to upload a file? Copy the skill text instead, paste it into a new chat and say “save this as a skill”. It gives the same answers, but it's slower every time you use it: with no script files, Claude has to rebuild the script at the start of each new chat, which adds some seconds to your first check in every conversation.
Create a Project called “Levelz”, paste these into its instructions, and pin it so it's always one tap away.
Open a chat and say “check my levelz.”
In ChatGPT, open GPTs → Create → Configure. Name it Levelz.
They include the educator voice, safety rules and response format.
Under Actions, choose Create new action, keep Authentication as None (your read token is already in the schema), and paste the schema.
Save it as “Only me”, since the schema holds your token. Then say “check my levelz.”
Levelz is for forward-looking people who are comfortable trying something new. It hasn't been properly tested. It's one person's DIY setup, shared in the #WeAreNotWaiting spirit, so you use it at your own risk.