AI for Astrophysics Summer School
A 2.5-day hands-on school on using AI assistants and coding agents well in astronomy research.
- Why this school
- Who it is for
- Programme at a glance
- Practicalities
- Before you arrive
- Reading the programme
- Talks
- Hands-on
- Habits
- What you take home
Why this school
AI assistants have moved from novelty to everyday research tools. They can read a codebase, run analyses, search the literature and draft text. They can also invent citations, introduce silent bugs and produce results that look right but are not. Most astronomers are learning to use them by trial and error, on their own.
This school gives you a structured, hands-on start: how these models work, how to direct them, and above all how to check their work.
Our organising idea
Good use of AI is mostly good research-software practice (version control, clear specifications, tests, review and documentation) combined with the scientific judgement only you can bring. These habits make an agent’s output checkable, and agents make the habits cheaper to adopt. The habits run through every talk, lab and project.
Who it is for
Thirty astronomers, mainly PhD students to senior postdocs; Masters students, research software engineers and academic staff are also welcome to apply. You need basic command-line skills (moving between folders, running commands). No experience with AI tools is required. Places are balanced across experience, with priority to those who will gain the most; experienced users get optional extra talks, stretch goals and a mentoring role.
What you will be able to do afterwards
- Explain how large language models work, and why they fail in the ways they do
- Direct an AI coding agent through a real analysis and review every change it makes
- Use git, tests and clear specifications to keep an agent’s work checkable
- Use AI for literature work and writing without importing its errors
- Set out how you, and your research group, will use AI responsibly
Programme at a glance
Draft programme. The school programme is still coming together and is subject to change: sessions, times and speakers may move as plans are finalised. Items marked TBC are still to be confirmed.
Mornings are talks; afternoons are hands-on. Day 1 ends with a guided lab; Days 2 and 3 are team projects, finishing with a show and tell on the afternoon of Day 3.
| Time | Day 1 · Mon 8 FebUnderstand and use | Day 2 · Tue 9 FebDirect and verify | Day 3 · Wed 10 Feb (half day)Finish and share |
|---|---|---|---|
| 09:00–10:30 |
Welcome T1 How LLMs work Talk–try |
T4 AI as a research assistant Talk–try |
Stand-up, then Project sprint 3 |
| 10:30–11:00 | Morning tea | ||
| 11:00–12:30 |
T2 From chatbots to agents Talk–try with live demo |
T5 Trust but verify Talk–try |
Project sprint 3 continues Code freeze at 12:30 |
| 12:30–13:30 | Lunch | ||
| 13:30–15:30 |
T3 Engineering for agents Talk–try, 45 min then the guided Lab in pairs |
T6 Responsible use 45 min then stand-up and Project sprint 1 |
13:30–14:30 Show and tell 14:30 Close |
| 15:30–16:00 | Afternoon tea | ||
| 16:00–17:30 |
Lab wrap-up Project pitch fair and team formation |
Project sprint 2 Optional alongside: T7 Advanced agentic workflows End-of-day failure log |
|
| Evening | Welcome social |
Optional: T8 The research frontier Lightning talks and unconference |
|
- Talk
- Hands-on: lab or project sprint
- Break
New to terms such as talk–try or project sprint? See Reading the programme. Dots show the range of experience each session is pitched at: , , . Project sprint time totals about 6 hours, so most teams aim for Silver; Gold is there for those who want it.
Practicalities
| Item | Details |
|---|---|
| Venue and travel | Swinburne Virtual Universe, room ATC103, Hawthorn campus. See local information for getting there and where to stay. |
| Bring | A laptop (macOS, Linux, or Windows with WSL) and its charger. |
| AI tools | Every school participant receives a one-month Claude Pro subscription, which includes both the Claude chat assistant and Claude Code, the command-line coding agent, for one month starting about a week before the school. Concepts are taught so they transfer between tools. Use comes with limits that reset every few hours, and heavy agent use can reach them; we will show you how to work within them, and tutors have back-up access. |
| Open science | All materials are open source and all project work happens in public repositories. A data policy will be shared before the school; do not put data into AI tools that you could not share publicly. |
| Support | Tutors at roughly one per 6–8 participants, plus a chat channel open from Day 0. |
| Speakers and tutors | To be announced |
| Workshop | The school is followed by the 21st ANITA Science Workshop, Thu 11 – Fri 12 February. One registration covers the whole week; see the workshop page. |
| Code of conduct | All participants are expected to follow the ASA code of conduct. |
| Contact | dcroton@swin.edu.au |
Apply by Tuesday 1 December 2026 (places confirmed Monday 7 December).
Before you arrive (Day 0)
Setup problems should never eat into lab time, so a little preparation happens online in the week before the school. Your one-month Claude Pro subscription (see Practicalities) starts about a week before the school.
- Complete the short pre-school survey on your experience and goals, so we can form balanced teams.
- Pre-work, about 3 hours: a short free introductory course on the coding agent we will use (details to follow); GitHub’s “Introduction to GitHub” course (or equivalent git refresher).
- Accounts: a GitHub account and a free NASA ADS API token.
- Install the coding agent on your laptop (macOS, Linux, or Windows with WSL); instructions will follow.
- Join the 60-minute online setup clinic about one week before the school.
- Optional: to bring your own project, tick “Bring your own project” on the registration form and add a one-paragraph pitch.
Reading the programme
| Symbol or term | Meaning |
|---|---|
| Little or no experience with AI tools. | |
| You have used chat assistants and perhaps tried a coding agent, but would not call yourself confident. | |
| You use AI coding agents regularly in your research. | |
| Ranges | Each session shows the range of experience it is pitched at: , , . Most sessions span more than one level, so wherever you are on the scale there is something for you. |
| Talk–try | A talk that stops every 15–20 minutes for about 10 minutes of hands-on work with an AI tool, guided by a prompt card, then a 2-minute debrief. You need your laptop open. |
| Guided lab | A step-by-step exercise in a prepared repository. The AI agent acts as your tutor, and each checkpoint has a test that tells you whether you are on track. |
| Project sprint | A block of team project time with tutors and speakers on hand. The deeper, hands-on part of the school, where the Bronze, Silver and Gold tiers let each team pitch its work to its experience. |
| Bronze / Silver / Gold | Project tiers. Teams decide how far to climb. Bronze is achievable by a newcomer; Silver extends it; Gold is a stretch. |
Talks
Six core talks for everyone, plus two optional advanced sessions. Talk–try sessions need your laptop open. Speakers: to be announced.
| Talk | Title | When | Pitched at | Format |
|---|---|---|---|---|
| T1 | How LLMs work: from tokens to transformers | Day 1, 09:15 | Talk–try | |
| T2 | From chatbots to agents | Day 1, 11:00 | Talk–try + demo | |
| T3 | Engineering for agents | Day 1, 13:30 | Talk–try | |
| T4 | AI as a research assistant | Day 2, 09:00 | Talk–try | |
| T5 | Trust but verify | Day 2, 11:00 | Talk–try | |
| T6 | Responsible use: policy, authorship and data | Day 2, 13:30 | Talk + discussion | |
| T7 | Advanced agentic workflows (optional) | Day 2, 16:00 | Talk–try | |
| T8 | The research frontier (optional) | Day 2, evening | Talk + Q&A |
T1 · How LLMs work: from tokens to transformers
What a large language model actually is: tokens, embeddings, attention and next-token prediction, and how training and post-training turn a text predictor into an assistant. Explains why these models are fluent but not grounded, the root of most of their failures.
Ask the assistant to count the letters in a word and explain any failure through tokenisation; ask for a paper on a niche topic and check whether it exists.
Connects to: T5; the lab
T2 · From chatbots to agents
How a chat model becomes an agent that reads files, runs code and edits a repository: tool use, the agent loop and context. A tour of a modern coding agent (permissions, planning mode, instruction files, skills, subagents and connections to outside tools), ending with a live demo of a real astronomy task, mistakes included.
Give the agent a small task in planning mode and read the plan before approving; add a three-line instruction file and repeat.
Connects to: the lab; all projects
T3 · Engineering for agents
The school’s organising idea: good AI use is mostly good software engineering. Version control, small steps, clear specifications, tests, reviewing diffs and documentation make an agent’s work checkable; working safely (permissions, secrets, what not to paste) comes with them.
Ask for a change with no specification, then again with acceptance criteria and a failing test, and compare the diffs; break something, then recover it with git.
Connects to: the lab; all projects; every project’s definition of done
T4 · AI as a research assistant
Using AI assistants for literature search, synthesis, drafting, editing and critiquing proposals, alongside ADS and arXiv, and how to judge the answers you get. Covers where it helps, where it invents, and what journals and peer-review rules allow.
Ask for five key references on a topic, verify each against ADS, and score the hit rate.
Connects to: T5; all projects
T5 · Trust but verify
The failure modes that matter in research (invented citations, silent numerical bugs, plausible but unphysical results) and the habits that catch them: tests as specifications, cross-checks against known results, reproducibility, and logging how AI was used.
Give the agent a function with a planted unit-conversion bug; have it write tests first and see whether they catch it.
Connects to: all projects; the failure log
T6 · Responsible use: policy, authorship and data
How research groups are setting AI policies, what journals expect you to disclose, authorship and data stewardship, energy cost, and how to keep building your own skills, especially as a student, while using AI.
Draft a one-paragraph AI policy for your own group.
Connects to: the group AI policy you take home
T7 · Advanced agentic workflows (optional)
For experienced users: subagents, parallel work in separate branches, unattended runs, hooks, and writing your own skills and tool connections for astronomy archives.
Launch two subagents on independent tasks and compare their diffs before merging.
Connects to: Gold-tier work in all projects
T8 · The research frontier (optional)
Where research is heading: astronomy-specific language models, agents that analyse data end to end, how they perform on hard problems, and the cost of running them at survey scale.
Connects to: discussion for anyone curious
Hands-on: lab and projects
Day 1 afternoon is a guided lab that everyone does. Days 2 and 3 are team projects: choose a theme from the menu or bring your own.
Lab · Your first agentic analysis (guided lab)
Day 1, 14:15–17:30 · Pairs · · Connects to T2, T3, T5
A prepared repository walks you through a short astronomy analysis on public survey data. You direct an AI coding agent; the repository keeps you on track, with the agent set up as your tutor and a test at every checkpoint. Every step ends with a reviewed diff and a commit, so your git history becomes your lab record. Core steps suit newcomers, and extension steps give more experienced pairs plenty to do.
You will learn to:
- Understand what a coding agent is: a model, its harness and the tools it can call
- Give a coding agent clear instructions and review every change it makes
- Check the agent’s claims against the code and against known results
- Use git as a safety net
- Write a skill that teaches the agent a task you repeat
- Write a hook that automatically checks the agent’s work
Projects (Days 2–3)
Teams of 3–4 · · Project sprints 1–3
Projects run in mixed-experience teams of 3–4 (or in pairs), formed at the pitch fair on Day 1 using your survey answers, with about 6 hours of project sprint time across Project sprints 1–3. Every project has Bronze, Silver and Gold tiers, so a newcomer and an expert on the same team both have real work. All work happens in public repositories. Prefer your own research problem? Bring it: tick “Bring your own project” on the registration form and add a one-paragraph pitch, and a tutor will help your team set Bronze, Silver and Gold goals at the pitch fair. Everything is done in the open, so choose code and data you can share.
The project menu will be announced closer to the school tbc. The themes below show the kinds of project on offer.
What the tiers mean
| Tier | What it means | Illustration: reproduce a published result |
|---|---|---|
| Bronze | Achievable by a newcomer. | Reproduce one figure from the public data alone. |
| Silver | Extends Bronze; where most teams aim. | Reproduce the headline number, with uncertainties. |
| Gold | A stretch for those who want it. | Add a sensitivity test or an extension the paper did not do, in a fully reproducible pipeline. |
Example project themes
Modernise a legacy research code
Get an old code running, pin down its behaviour with tests, then fix, port or speed it up, checking the physics and not just the old outputs.
Good for: code, testing, performance
Reproduce a published result
Reproduce a figure from a published paper using only the public data and your own analysis, then explain any differences.
Good for: analysis, reproducibility
From archive to science result
Go from a public archive query to a science result, checking each step against known answers.
Good for: data analysis
Verified literature review
Use an AI assistant to review a live debate in astronomy, then verify every claim and citation it makes.
Good for: literature, evaluation
Build a toolkit for your group
Build the agent instruction files, skills and tools your research group would actually use, and share them with the cohort.
Good for: skills and tools for your group
AI as a proposal reviewer
Explore where AI critique of a mock proposal helps, and where it is unreliable or inconsistent.
Good for: writing, peer review
Bring your own
Work on your own research code or problem, set up with a one-paragraph pitch and tiered goals agreed with a tutor.
Good for: your own research problem
Definition of done
Every project, at any tier, is checked against this list at the show and tell:
- A public repository with a meaningful commit history
- A README and agent instruction file that let a newcomer, or an agent, pick the project up
- Passing tests, including at least one written before a change
- A pinned environment and a one-command way to run it
- A failure log: what the AI got wrong, and which habit caught it
Silver adds automatic testing on every push; Gold adds packaging and a release.
A day in the project sprints
- Stand-up (10 min) at the start of each project day: your goal, your tier, your blocker.
- Help on hand: tutors circulate and speakers stay for office hours. Put up a green sticky note when all is well, pink when you need help.
- Failure log: each team records what the AI got wrong and how it was caught. The best go on a shared board.
- Show and tell (Day 3, 13:30): 4 minutes per team, including one slide titled “What the AI got wrong”.
The habits that run through everything
| Habit | Why it matters when an agent writes code | Where you will practise it |
|---|---|---|
| Version control, small steps | Every change can be reviewed and undone; small changes hide fewer errors | The lab (commit every step); all projects |
| Clear specifications | A goal with acceptance criteria is the best prompt | T2, T3; project tiers |
| Tests | The contract with the agent; write them before changing code | T5; the lab; all projects |
| Reviewing diffs | The agent can review too, but a person signs off | Project sprints; failure log |
| Documentation as context | README and agent instruction files (e.g. AGENTS.md) are what the agent reads first | The lab; all projects |
| Reproducible environments | Anyone, human or agent, can rerun your result | Pre-work; all projects |
| Automation and profiling | Tests on every push catch regressions; measure before optimising | Silver and Gold tiers |
Passing tests are not the same as correct physics. Every project also asks the team to check results against independent expectations: known values, limiting cases, published figures.
What you take home
- Your team’s public repository, built to the definition of done
- A shared library of the cohort’s skills and agent instruction files
- A draft AI policy for your own research group
- About three more weeks of AI tool access after the school, with an online check-in before it ends
- A cohort channel to keep working together, with the option of a joint write-up of the school’s projects
Join us at Swinburne
Places are limited to 30 and are balanced across experience. Apply by Tuesday 1 December 2026; later applications join a waitlist. One registration form covers the whole week, so tell us which parts you will attend.