Retext
AiderCopilotFor ML Engineers

Aider Prompts for ML Engineers

ML engineers need help with training and evaluation code, data pipelines, debugging shape and dtype errors, and making models reliable and fast in production. These Aider prompts start from that: fill in your situation once and reuse it. The worked example below is a debugging task, one of the most common for ml engineers.

Installs as a snippet — type ;ml-aider to paste this prompt anywhere on your Mac.
Prompt Template
I'm a ml engineer and need help with [task, e.g. model training and tuning].

What I have:
- [the material involved: code, draft, notes or data]

What a good result looks like:
- [the outcome and its format]

Ask me one clarifying question if something important is missing, then give the result.

About my work
- Framework and hardware: [e.g. PyTorch 2.3, single GPU with 24 GB]
- Model and data: [e.g. fine-tuning a text classifier, 200k labeled examples, 5 classes]
- Production constraints: [e.g. p95 latency under 100 ms, CPU inference only]

Context
- Files involved: [list the files, or reference them in the chat]
- How I build and test: [e.g. npm test]

How to use this prompt

  1. 1

    Open Aider

    Start aider in your repository and paste the prompt at the aider prompt.

  2. 2

    Give it context

    Add the files it should edit with /add before you send the prompt. Aider only edits files you have added.

  3. 3

    Review and iterate

    Aider commits each change to git, so review with git diff and use /undo if a change is wrong.

Working in Aider

  • •Only /add the files that need changes. Extra files cost context and invite unrelated edits.
  • •Because every change is a git commit, you can keep requests small and roll back cheaply.
  • •Paste test output or error messages directly; aider works well in a fix, run, paste-error loop.

Example task: Debugging

What you give it

Symptom: add_tag returns tags from earlier calls. Calling add_tag("a") then add_tag("b") gives ["a", "b"] on the second call instead of ["b"].
Environment: Python 3.11, local script.

def add_tag(tag, tags=[]):
    tags.append(tag)
    return tags

What changed: nothing, it has always done this.

What a good answer looks like (shortened)

1. Mutable default argument (very likely). The list [] is created once, when the function is defined, and every call without a tags argument reuses that same list.
   Check: print(id(tags)) inside the function. If the id is the same on every call, this is the cause.

2. A caller is passing the same list in each time (less likely, since the calls you showed pass only one argument).
   Check: search for calls to add_tag that pass a second argument.

Run check 1 first. If confirmed, the fix is:
def add_tag(tag, tags=None):
    if tags is None:
        tags = []
    tags.append(tag)
    return tags

Tips for better results

  • •Include tensor shapes and the full traceback when debugging. Most training errors are shape or device mismatches that are obvious once the AI sees the actual dimensions.
  • •Add the files it should edit with /add before you send the prompt. Aider only edits files you have added.
  • •Check this before you use the output: check evaluation code for leakage between train and test data and confirm reported metrics are computed the way you think, since a small bug can make a model look far better than it is.

Check before you use it: Check evaluation code for leakage between train and test data and confirm reported metrics are computed the way you think, since a small bug can make a model look far better than it is.

Get the free prompt pack

Our best coding prompts in one Markdown file, delivered by email. The templates here stay free either way — the pack just saves you the copying.

Variations

Task

Debugging: Replace the first line with: Help me debug a problem. I want to find the cause, not just patch the symptom.

Intermittent bug

Add this line to Symptom: It only happens [e.g. 1 in 20 requests, under load, after midnight UTC]. Focus on timing, concurrency and caching causes.

Second task

Change the task line to: [ML pipeline development].

Save it as a Retext snippet

Retext is a free text expander for Mac. Save this prompt once, then type ;ml-aider in any app (ChatGPT, Claude, your editor, a browser tab) and the full prompt is pasted in.

  1. Click Add to Retext. The prompt installs as a snippet with the trigger ;ml-aider.
  2. Change the trigger in Retext if you prefer a different one.
  3. Type the trigger wherever you want the prompt, then fill in the brackets.

Frequently asked questions

Can AI help tune hyperparameters?

It can suggest sensible ranges and a search strategy based on your model and data size. The actual best values still have to come from experiments on your validation set.

Why didn't aider change the file I meant?

Aider edits only files added to the chat. Run /add with that file, then resend the prompt.

How do I undo a bad edit?

Use /undo in aider to revert its last commit, or use git as usual, since each change is committed.

How do I reuse this prompt without copying it every time?

Click Add to Retext to save it as a snippet. Then type ;ml-aider in any Mac app, including your AI tool, and the full prompt is pasted in.

Related Templates

Get the free prompt pack

Our best coding prompts in one Markdown file, delivered by email. The templates here stay free either way — the pack just saves you the copying.