Marauder Bot

There are two ways to use AI on LeetCode-style problems. One makes you dramatically better; the other manufactures the feeling of progress while your actual skill flatlines. The difference is not the tool — it's the workflow. Here is the one that works.

The failure mode first

The trap: read problem → ask AI for solution → read solution → nod → "learned it" → next problem. Fifty problems later, you've built a superpower for recognizing solutions and no ability to produce them. Recognition feels identical to competence right up until an empty editor and a running clock, which is exactly where interviews happen.

The principle that prevents it: AI enters after the attempt, never before. Everything below is that principle turned into a routine.

The workflow

1. Attempt cold, under a timer

Twenty-five minutes for a medium, in one sitting, no lookups. Struggling is the training — retrieval under mild pressure is what builds the pattern memory you'll need. Reaching the time limit without a solution is a fine outcome; you've now loaded the problem into your head, which makes the review stick.

2. Stuck? Take graduated hints, not answers

Ask your AI for the smallest useful push, escalating only as needed:

"Give me a nudge — no approach, no code." (e.g. "what if the array were sorted?")
"Name the pattern." ("this is a sliding-window problem")
"Outline the approach in three steps, still no code."
Full solution — only after a real attempt at each rung.

Every rung you don't descend is skill kept. Chatbots follow these instructions well if you state them explicitly.

3. Review against a complete answer — all four parts

After the attempt (solved or not), study a full worked answer: working code, test cases, time complexity, space complexity. Compare it against yours: Where did the approaches diverge? Which edge cases did their tests cover that yours missed? Is your complexity actually what you thought?

This is the step Marauder Bot compresses. It reads the problem page directly in Chrome — no copy-paste, and it works on sites that disable text selection — and returns exactly that four-part answer for every question it finds (see the live demo). With three switchable models (GPT-5.6 Luna, Claude Opus 4.8, Gemini 3.5 Flash), a shaky answer gets a second opinion in one click. What took an evening of forum-thread archaeology takes seconds, which is what makes the review habit sustainable.

4. Interrogate until the "why" is yours

The worked answer tells you what; make the AI earn the why: "Why a heap and not sorting?" "When would this two-pointer approach break?" "Show me the naive solution and walk the optimization." A solution you can derive beats ten you've memorized — variations of a pattern show up far more often than exact repeats.

5. Re-solve from scratch, days later

The step everyone skips and the one that cements everything: three or so days after learning a problem's pattern, solve it again cold. If you can't, it was recognition, not competence — review again, re-queue it. (The JSON export of your analyses is handy for keeping this queue in a spreadsheet or flashcard deck.)

Weekly structure

Practice in patterns, not random problems: a week on two-pointers, a week on BFS/DFS, and so on — three or four new problems a day through steps 1–4, plus two or three re-solves from earlier in the week. Add a weekly timed mixed session under assessment conditions to train pacing alongside patterns.

The one-sentence version

Attempt cold, hint minimally, review completely, interrogate the why, re-solve later — AI makes every step faster and none of them skippable. Run the loop for a few weeks and the difference is obvious in the editor, which is the only place it counts. If the review step is your bottleneck, the free trial covers a week of practice sessions: 15 page analyses over 7 days.