8 min read
This article is currently shown in English. A translation is in progress.
LoL2LLM ships five built-in personas (Coach, Roast, Praise, Caster, Analyst) and copies the right system prompt along with your match JSON. Pasting and hitting enter works, but a single extra line of context after the JSON typically doubles the usefulness of the response. Here's when to use each persona, and the four prompt additions that make all of them sharper.
Coach: the default. Balanced criticism with concrete improvement steps. Works on both wins and losses; if you're unsure, start here.
Roast: for games you already know you played badly. Brutal, no-mercy critique. Useful when you need a kick to drive change — not when you're tilted.
Praise: a mental-reset tool after a tilting loss streak. Defends you and credits team-gacha for the loss. You learn nothing — that's the point. You come back ready to focus next game.
Caster: e-sports shoutcaster style. Best for sharing a game with friends or a stream chat — the dramatized read becomes a fun copy-paste comment, not a coaching note.
Analyst: cold, numeric evaluation. CS deltas, gold efficiency, damage per gold. Closest to what a paid pro coach's spreadsheet review feels like.
Persona prompts are intentionally generic, so the AI returns generic feedback. One sharper line at the end of the paste:
I lost CS badly to my laner this game. Restrict the analysis to the lane phase and tell me when and why my decisions probably went wrong.
Without context, the AI defaults to evaluating against pro play. Saying "I'm a Gold III jungle main" reshapes the advice into things actually executable at that level.
I'm a Gold III jungle main. Skip pro-level macro suggestions; only give advice that's realistically executable at this rank.
Left alone, an LLM will produce 10–15 items of feedback. You can't fix all of them next game, and now you have to do the prioritization work yourself. Asking for the top three forces the model to rank, and gives you a focused practice list.
Limit feedback to the three highest-impact issues. For each, give one sentence on why it mattered and one sentence on what to do differently next game.
Paste your last five exports in sequence and ask "what's the recurring pattern across these games?". Single-match analysis can't catch repeating habits. You'll often hear something specific you didn't notice yourself — "you burn flash defensively before 10 minutes in 4 of 5 games" type observations.
Both are competent; their strengths differ. ChatGPT (GPT-4o, o1) tends to produce cleaner bulleted action plans. Claude (Sonnet, Opus) is better at long-form, narrative reads of complex teamfights and macro mistakes — it'll explain why a rotation was wrong rather than just naming it. For Roast, ChatGPT bites harder. For dense teamfight breakdowns, Claude reads more like a coach. Pasting the same JSON to both and comparing is a legitimate workflow.
Current meta and patch-specific tier lists are unreliable from an LLM — the knowledge cutoff is months old and bans/buffs move fast. For "is this champ strong this patch" questions, go to op.gg or u.gg. What an LLM is excellent at is analysis bounded by the JSON you handed it: lane decisions, item paths in this matchup, fight positioning given these comps. Stay inside the JSON and the answers stay sharp.