Mirror of forum.nim-lang.org

13174 :: Nim Coding with AI: Best Practices

• martinc (orginal) [2025-07-06T14:31:33+02:00] view original

  Hey everyone! I recently began trying out the AI built into the Cursor editor with some Nim projects I've been working on (saw suggestion about using Cursor from @Araq a while back) I wanted to ask if anyone has some best practices/suggestions about how to get the best results from the AI when using the Agent mode. The issue I'm running into is with Nim (more often than a more mainstream language like Python) the IA is hallucinating features of the language that don't exist. For example:

  
  var s: seq[int]
  s.add(42)        # append “42” as new last element
  echo s[^]        # 42
  Run

  and

  
  var i = 0
  
  # do { i += 1; echo i; } while i < 3;
  repeat
    i += 1
    echo i
  until i >= 3
  Run

  These are contrived examples but I was able to get the AI to hallucinate things like the [^] last element index, as well as repeat and until keywords which don't exist! So if anyone out there is able to get Cursor working well on creating larger code blocks I'm interested in your approach.

  Thanks!

---

• UxDnz0 (orginal) [2025-07-06T15:21:26+02:00] view original

  2 ways I do: Implement its hallucinations if they're pithy. Use AI to vibe grasp my code base but never to write the code, just typing it out is great for my memory. Vibe code < Vibe understand(&test properly)

---

• elcritch (orginal) [2025-07-06T23:03:46+02:00] view original

  Odd I’ve had Cursor hallucinate some things but usually not basic syntax like that. More imagining library functions.

  Which agent are you using? I’d recommend Claude 3 or 4. Who knows what Cursors default model uses. I also find Gemini capable.

  I find Cursors $20/mo the best deal as well. I haven’t hit limits on the above models yet.

---

• elcritch (orginal) [2025-07-06T23:06:13+02:00] view original

  Also in case you’re not aware, the inline edit function hallucinates much more. It’s a very simple model.

  If you want to generate bigger bits of code use the chat box. That’s where you can set the model.

---

• elcritch (orginal) [2025-07-06T23:38:13+02:00] view original

  One final note, YMMV. Prompts take a bit of art to learn and vary between model, even model sizes. With the smaller Gemini-2.5-flash model adding idiomatic Nim actually made it generate incorrect code while Gemini-2.5-pro wrote idiomatic Nim either way.

---

• ggibson (orginal) [2025-07-07T15:33:41+02:00] view original

  A few years ago models struggled a lot. I would use this trick of pasting in a portion of the manual or learn X in Y min, then asking for a concise description of in which ways Nim syntax is different than python (correct or add manually as necessary). Then I stick that in a part of every prompt going forward, what people now call a rules file in cursor. So, I didn't care if it had Never seen nim code before. This helped a lot.

---

• cmc (orginal) [2025-07-07T17:56:57+02:00] view original

  If you check out a C repository, I find claude code rather effective at generating a Nim wrapper. I was also able to translate C test code to use the wrapper and have the test output match the original. This seems to be a sweet spot for AI because both activities are tedious and require knowledge, but not much depth of thought.

---