How do I keep my coding skills while using AI?

Keep one decision yours before asking for help

Choose one small piece of real work each day and make a prediction before you open an assistant. Write what the function should return, what it may change, and one input that could break it. Then use your normal tools to check. This gives you a concrete record of your reasoning while leaving room for AI in the rest of your workflow.

The useful question is which actions you can still perform independently. Can you reconstruct a loop? Follow a value through three calls? Explain why a proposed fix satisfies the requirement? Treat these as separate things to practise. Finishing a task with an assistant does not, by itself, tell you which of them you used.

Try it: what changes outside this function?

A helper should return the two smallest numeric IDs while leaving the caller's queue unchanged. Before running this JavaScript, write both output lines. Then decide whether the helper meets the requirement. Give yourself time to trace each operation; this is an exercise, not a timed assessment.

const queue = [17, 3, 11];

function nextTwo(ids) {
  return ids.sort((a, b) => a - b).slice(0, 2);
}

console.log(JSON.stringify(nextTwo(queue)));
console.log(JSON.stringify(queue));

function nextTwo(ids) {
  return [...ids].sort((a, b) => a - b).slice(0, 2);
}

Use a ten-minute practice slot

Here is a starting schedule, not a validated dose: spend two minutes choosing a small requirement and predicting the result, four minutes implementing or tracing it without hints, and four minutes checking and explaining the mismatch. If the problem will not fit, shrink it. One branch or one collection operation is enough.

After checking, write a short note about the decision you missed. For this example, the note might be: find the first operation that can change shared state. Tomorrow, try that question on a fresh example with different data. Repeating the exact output tests whether you remember this page; a changed example asks you to use the rule.

Measure the part you want to retain

Keep three observations: what you attempted before help, where your explanation stopped, and what evidence changed your answer. If syntax is the obstacle, reconstruct a short function. If the result surprises you, trace values. If the code looks right until an unusual input arrives, practise finding counterexamples.

A difficult session is a reason to inspect your approach, not proof that AI caused a decline. Task difficulty, familiarity and the conditions of the attempt can all change what you observe. To compare attempts, keep the language and assistance conditions consistent and use fresh problems.

Where Unrust fits into this workflow

Unrust's free diagnostic samples six dimensions with AI and hints closed: recall, tracing, comprehension, debugging, idiom and AI-slop detection. Use the profile to choose a practice target rather than treating one headline as a verdict on your ability. A single sitting describes that sitting; it cannot establish that a tool caused skill loss.

After that baseline, the adaptive practice loop lets you work on individual skills and revisit them over time. Keep building real software alongside it. Short drills isolate a reasoning decision; a project adds the requirements, dependencies and tradeoffs those drills cannot fully reproduce.