Every time a CCD reads out an image, some electrons get caught in defects in the silicon and are released a moment later, smearing faint trails behind stars and galaxies. This is Charge Transfer Inefficiency (CTI), and PyAutoCTI measures and corrects it.
This repository is the PyAutoCTI Assistant: an AI assistant which lets you use natural language to calibrate and correct CTI in CCD imaging.
It's a young assistant, built from the more mature PyAutoLens Assistant. The basics work today, and its CTI know-how grows as people use it.
The assistant runs inside an AI coding agent, a tool that can read this repository, run Python on your computer and look at the results. Pick your agent: Claude Code or Codex are recommended, and GitHub Copilot CLI or OpenCode are free AI options if you want to try it without paying.
Once it is installed, download this repository and open your agent inside it, then paste in the starting prompt
below. The assistant installs PyAutoCTI for you after your first prompt. Its arctic dependency needs a compiler
and the GSL library, and skills/ac_setup_environment.md has the details if
anything goes wrong.
git clone https://github.com/PyAutoLabs/autocti_assistant.git
cd autocti_assistant
claude # or: codexIf you've not used the PyAutoCTI Assistant before, use the prompt below. It explains CTI, runs a small calibration and shows you the result.
Starting prompt for Claude Code, Codex or other AI coding agent
I'm new to PyAutoCTI. Explain what charge transfer inefficiency is and how a calibration works, then run the
1D calibration demonstration in scripts/demonstrations/demo_1_calibrate_1d.py, show me the fit, and check that
it recovers the input trap density and release timescale.
If you're new to CCD detectors, start your prompt with Teacher mode. and the assistant will explain the physics as
it goes. You can also ask it to plan a calibration and talk through the trap model with you before anything runs.
- Four demonstrations you can run in
scripts/demonstrations/: calibrate a 1D dataset, calibrate a 2D charge-injection image, correct CTI, and load the results of a fit. Each one checks its own answer, so a clean run shows everything is working. - A reference wiki in
wiki/core/, covering what CTI is, how traps behave, the arctic algorithm and how to plan a calibration. - A literature wiki in
wiki/literature/, with a checked bibliography. - Benchmarks in
benchmarks/. No CTI benchmark has been written or run yet, so there are no performance claims here.
The list of skills the assistant has today is in skills/README.md, and what's planned next is
in PENDING.md.
When you start a real scientific study, ask the assistant to set up a science project. This is a separate repository that holds everything for that study:
- Data. The datasets and configuration for the study.
- Analysis. The scripts the assistant writes and the results they produce.
- History. A record of your work with the assistant, so you can pick up where you left off.
To start one, just ask:
Example Natural Language Prompt for Claude Code, Codex or other AI coding agent
Start a science project for my Euclid VIS CTI calibration.
The assistant ships agent instructions and reference material derived from the public PyAuto* repositories; the underlying libraries are released under their own licenses (see each repo).
Note for AI coding agents
If a prompt sent you here and you are not already inside this repository, clone it
(git clone https://github.com/PyAutoLabs/autocti_assistant.git && cd autocti_assistant), then read AGENTS.md
in full and follow it. Science projects are set up by skills/start-new-project.md.