Don't underestimate what Claude can do with your NINA Advanced Sequencer JSON

andrea tasselli · Mar 24, 2026, 10:33 PM

All of which can be done with a modicum of brain power and no AI.

I like this fighting spirit but I also like jumping the learning curve.

Spacey · Mar 24, 2026 at 11:18 PM

andrea tasselli · Mar 24, 2026, 10:33 PM

All of which can be done with a modicum of brain power and no AI.

I like this fighting spirit but I also like jumping the learning curve.

No reason not to do both.

Frank "Voloire" · Mar 24, 2026, 10:15 PM

Been using Claude (Anthropic's AI) as a hands-on co-pilot across my entire AP workflow — not just for generic questions, but with direct access to my files, FITS headers, and sequencer JSON. Here's what a single session looked like.

It started by reading my FITS headers and correctly identifying my target as LBN 239 & LBN 251 (IC 1318 / Gamma Cygni complex) from raw RA/Dec coordinates — I had the wrong name in my notes.
It then parsed 164 light frames across 5 nights, built a per-night airmass table showing exactly when each session dropped below 2.0 and when the best transparency window hit (consistently 04:30–05:15 local).

From that airmass data it proposed a smarter filter order: HA first (strongest signal, tolerates high airmass at session start), SII in the middle, OIII always last — consistently shot in the best transparency window every night.
Obvious in retrospect, easy to miss when you're just clicking through NINA's loop UI.

Then it rewrote my entire multi-target sequence JSON from scratch — four targets (Sadr SHO+RGB, M65, M86, M13 LRGB), removed all loop/iteration structures, replaced them with clean sequential SmartExposure blocks, calculated exactly how many additional frames each filter still needed to reach balance against existing data..

Finally — and this is the part I care about most — it added park-on-failure safety logic: if plate solve or autofocus fails (assuming: clouds, bad seeing), the mount parks immediately rather than keeping tracking while waiting for the next target to rise. Real pier-crash prevention, not just a nice-to-have.

Three JSON versions, zero manual editing, all from plain conversation.

Ridiculous.

By using AI to help tune your sequence you’ve been exposed to a better way of doing things. I am sure you would have figured out on the changes on your own, but not quite so fast. You’ll now start evolving your own knowledge from a higher point than before.

You’ve jumped the learning curve.

AI brings the gatekeeper out in people, people who feel threatened. :)

Spacey · Mar 25, 2026, 08:49 AM

Frank "Voloire" · Mar 24, 2026, 10:15 PM

Been using Claude (Anthropic's AI) as a hands-on co-pilot across my entire AP workflow — not just for generic questions, but with direct access to my files, FITS headers, and sequencer JSON. Here's what a single session looked like.

It started by reading my FITS headers and correctly identifying my target as LBN 239 & LBN 251 (IC 1318 / Gamma Cygni complex) from raw RA/Dec coordinates — I had the wrong name in my notes.
It then parsed 164 light frames across 5 nights, built a per-night airmass table showing exactly when each session dropped below 2.0 and when the best transparency window hit (consistently 04:30–05:15 local).

From that airmass data it proposed a smarter filter order: HA first (strongest signal, tolerates high airmass at session start), SII in the middle, OIII always last — consistently shot in the best transparency window every night.
Obvious in retrospect, easy to miss when you're just clicking through NINA's loop UI.

Then it rewrote my entire multi-target sequence JSON from scratch — four targets (Sadr SHO+RGB, M65, M86, M13 LRGB), removed all loop/iteration structures, replaced them with clean sequential SmartExposure blocks, calculated exactly how many additional frames each filter still needed to reach balance against existing data..

Finally — and this is the part I care about most — it added park-on-failure safety logic: if plate solve or autofocus fails (assuming: clouds, bad seeing), the mount parks immediately rather than keeping tracking while waiting for the next target to rise. Real pier-crash prevention, not just a nice-to-have.

Three JSON versions, zero manual editing, all from plain conversation.

Ridiculous.

By using AI to help tune your sequence you’ve been exposed to a better way of doing things. I am sure you would have figured out on the changes on your own, but not quite so fast. You’ll now start evolving your own knowledge from a higher point than before.

You’ve jumped the learning curve.

AI brings the gatekeeper out in people, people who feel threatened. :)

So... basically, using an LLM within the world of astrophotography — as far as I'm concerned — means extracting known patterns that millions of users have contributed over the years, entered through different systems. And through a process we could call maieutic, turning that into knowledge I can actually access.

So it's not a broad, polarizing philosophical debate about whether or not to use these tools — it's exactly that: using a tool. A tool built precisely to allow access to shared knowledge, knowledge that millions of people have poured in over years through various means.

The maieutic process happens through prompting. Obviously, you need domain knowledge and subject-matter competence to do it well.

So yes — keep improving, keep learning. Because otherwise, what are we even here for?

andrea tasselli · Mar 25, 2026 at 10:17 AM

AI did it for you, at most you copied and pasted. Next step: let Claude or whatever run the whole gizmo; that would save you quite a bit of learning curve.

There are many of us who quite literally download calibrated masters and don’t bother with NINA or telescopes or mounts at all.

Arun H:
There are many of us who quite literally download calibrated masters and don’t bother with NINA or telescopes or mounts at all.

Spacey · Mar 25, 2026 at 12:12 PM

I did not do the learning they did, I began with their ideas and took it from there. I jumped the learning curve. Haha. :)

I don’t think that qualifies as jumping the learning curve. You could similarly say that Isaac Newton skipped the learning curve because he used ideas and learnings from Kepler, Galileo and others.

These days, instructors check whether lab reports are AI generated. I would say if you used AI to generate a report, you definitely skipped the learning curve!

Frank "Voloire" · Mar 25, 2026, 09:54 AM

Spacey · Mar 25, 2026, 08:49 AM

Frank "Voloire" · Mar 24, 2026, 10:15 PM

Been using Claude (Anthropic's AI) as a hands-on co-pilot across my entire AP workflow — not just for generic questions, but with direct access to my files, FITS headers, and sequencer JSON. Here's what a single session looked like.

It started by reading my FITS headers and correctly identifying my target as LBN 239 & LBN 251 (IC 1318 / Gamma Cygni complex) from raw RA/Dec coordinates — I had the wrong name in my notes.
It then parsed 164 light frames across 5 nights, built a per-night airmass table showing exactly when each session dropped below 2.0 and when the best transparency window hit (consistently 04:30–05:15 local).

From that airmass data it proposed a smarter filter order: HA first (strongest signal, tolerates high airmass at session start), SII in the middle, OIII always last — consistently shot in the best transparency window every night.
Obvious in retrospect, easy to miss when you're just clicking through NINA's loop UI.

Then it rewrote my entire multi-target sequence JSON from scratch — four targets (Sadr SHO+RGB, M65, M86, M13 LRGB), removed all loop/iteration structures, replaced them with clean sequential SmartExposure blocks, calculated exactly how many additional frames each filter still needed to reach balance against existing data..

Finally — and this is the part I care about most — it added park-on-failure safety logic: if plate solve or autofocus fails (assuming: clouds, bad seeing), the mount parks immediately rather than keeping tracking while waiting for the next target to rise. Real pier-crash prevention, not just a nice-to-have.

Three JSON versions, zero manual editing, all from plain conversation.

Ridiculous.

By using AI to help tune your sequence you’ve been exposed to a better way of doing things. I am sure you would have figured out on the changes on your own, but not quite so fast. You’ll now start evolving your own knowledge from a higher point than before.

You’ve jumped the learning curve.

AI brings the gatekeeper out in people, people who feel threatened. :)

So... basically, using an LLM within the world of astrophotography — as far as I'm concerned — means extracting known patterns that millions of users have contributed over the years, entered through different systems. And through a process we could call maieutic, turning that into knowledge I can actually access.

So it's not a broad, polarizing philosophical debate about whether or not to use these tools — it's exactly that: using a tool. A tool built precisely to allow access to shared knowledge, knowledge that millions of people have poured in over years through various means.

The maieutic process happens through prompting. Obviously, you need domain knowledge and subject-matter competence to do it well.

So yes — keep improving, keep learning. Because otherwise, what are we even here for?

I think one of the problems with AI these days is the very fact that its information base is the internet. There’s a lot of out of date or just plain wrong information out there in the astrophotography world. The few times I’ve used it it took several rounds of me saying “ah, no. That’s incorrect” followed by the AI saying “You’re obsoletely right!” Kinda shakes my confidence in the output right now.

andrea tasselli · Mar 25, 2026, 10:17 AM

It means nothing at all. A strongly doubt there are "millions" of us in any way shape or form and you signally failed to learn something through the exercise of brain power which in itself is the whole point of the "learning curve". AI did it for you, at most you copied and pasted. Next step: let Claude or whatever run the whole gizmo; that would save you quite a bit of learning curve.

I don’t see much difference in using AI as a tool to potentially show you a better way of doing things than watching any of the various videos or websites showing new techniques. Some aren’t very good, and some are excellent. It’s up to you to try them and sort them out. AI sometimes produces garbage, but sometimes it exposes a nugget of information you never would have thought of on your own.

We tend to find a way of doing things and sticking with it until disrupted by new information from an external source that can lead to improvement. AI is just one of those external sources.

Learning curves?

Oh, like buying a go-to mount and letting it help you learn the Universe as it takes you from star to star?

The fact that we are all on Astrobin shows we are using computers instead of the Pony Express.

Me, Myself, and I welcome all the electronic help I find useful. There wasn’t much help from egotists. I had to dig in and figure out what various AI (Computers and Applications) were trying to show me. And troubleshoot my software when it got corrupted. Still do.

EAA was and still is to me: Electronically Assisted Astronomy. Go-To mounts, Guiding equipment, Electronic Cameras, and Computers, etc.

Personally, over the decades I’ve used Stellarium to locate starting with browsing the program, and now as my object choosing App to input into NINA’s Framing Wizard to setup where my mount slews to.

It is working smarter, not harder.

In many ways, these discussions remind me of the whole man versus computer or man versus machine discussions of the past. At the end of the day, Frank is right. These tools are increasingly capable and increasingly adopted and will very soon, if no already, be superior to what you can do on your own. Rather than fight it, embrace and adopt it. Even in my own, non software profession, I am surprised by the kind of information I can get from AI. As for them being wrong some of the time - what percent of time do you think some human you ask for help is wrong? How often are you wrong? How often is something you discover on the internet wrong?