9 min read
"Show the AI my loss and tell me what I did wrong" is the right instinct.
Most of what a coach does is read match data without ego and put the gap between your play and role expectations into words — which is close to what ChatGPT and Claude are best at.
But most players stop at "paste a screenshot, type ‘analyze this’," get a generic non-answer, and give up.
This guide walks through how to feed the data and structure the prompt so the AI produces concrete, coach-level feedback instead.
Pasting the post-game screenshot has three structural problems.
Not enough information: a screenshot only shows final stats. The timeline — when and in what order things happened — is missing, so the AI has to guess why the 14-minute death occurred.
Misreads: image recognition swaps digits and misreads icons. Read CS 187 as 167 and every downstream conclusion is wrong.
It invites hallucination: when the data is vague, the AI fills the gaps with plausible generalities. That is exactly where "ward more" and "watch your positioning" come from — advice that fits anyone and helps no one.
The fix is simple: hand it structured text data (JSON).
When the numbers are present as exact text, the AI can compare you to your lane opponent and flag anomalies without image-recognition error.
Ideally the prompt contains, as actual numbers:
Your champion/role and your lane opponent's (the comparison is the spine of the analysis)
KDA, CS/min, gold, damage share, and vision score
Ideally death timings and game length (an early death and a late death mean opposite things)
Transcribing all of that by hand is tedious, so LoL2LLM exports Riot's Match-v5 data as JSON trimmed to exactly these fields.
Search a Riot ID → pick a match → check the fields you want → copy. The [Paste JSON here] placeholder in the prompts below assumes this export.
A good prompt has three parts.
Give it a role (who it answers as), make the ask a concrete verb ("extract" / "compare," not "analyze"), and add constraints to narrow the output (count, audience, format).
Satisfy those three and the answer quality changes dramatically. Generic template:
You are a League of Legends coach specializing in my role. Below is my match data (JSON). # Task List exactly 3 improvements I can fix starting next game. For each, pair "the evidence in the data" with "the specific next action." # Constraints - No generalities (e.g. "ward more", "watch positioning"). Only points tied to THIS match's data. - I'm Platinum. Write for that level. [Paste JSON here]
The key is to explicitly ban generalities.
That single line makes it much harder for the AI to emit ungrounded advice and sharply reduces hallucination.
The "exactly 3" cap matters too — a hard limit forces it to lead with the highest-priority points.
Laners (top, mid, ADC) get sharper answers anchored on the matchup gap.
Build a table comparing my CS/min, gold, and damage share to my lane opponent's. Then identify the ONE lane-phase (0-14 min) metric I lost, and name two plausible play mistakes that caused it. [Paste JSON here]
Junglers respond well to "where did your resources go."
I'm the jungler. Compare my CS/min and kill participation to the average at my rank. Diagnose whether I lean too far toward farming or toward ganking, and suggest concrete clock times to add more of the opposite. [Paste JSON here]
Support — KDA is noise, so push it toward vision and damage soaked.
I'm the support. Ignore KDA entirely. Using vision score, control wards, and damage taken (a proxy for peel), judge whether I did my job this game. Name exactly one weak point and give a numeric improvement target for next game. [Paste JSON here]
A single game is swung by luck — team strength, a feeding matchup.
Hand over your last five games on the same champion and have it extract the weakness that repeats across games; that's the reproducible problem worth practicing.
Below are my last 5 games on the same champion. Identify 2 weaknesses that recur across games. Only surface patterns common to multiple games, not a one-off accident in a single match. [Paste multiple games' JSON here]
Don't let it confuse cause and correlation: the AI tends to shortcut "low CS → that's why you lost." Ask back "did this gap actually affect the result, or is it a symptom?" to raise precision.
It invents what isn't in the data: if it asserts things the JSON doesn't contain — vision contests, shotcalling — that's hallucination. Constrain it: "mark anything not in the data as speculation."
Don't over-trust its meta knowledge: the model has a knowledge cutoff. Verify item and rune optimums against official patch notes or a dedicated site.
(1) Hand over structured data (JSON), not a screenshot.
(2) Write the prompt in the role / ask / constraints pattern and ban generalities.
(3) Don't stop at one game — extract the weakness that repeats across several.
Do just these three and AI coaching turns from "a bland impression" into "a concrete piece of homework you can try next game."
LoL2LLM handles the first step of that workflow — getting data worth analyzing ready in one click.
A field-by-field walkthrough of the JSON LoL2LLM hands to ChatGPT or Claude. Knowing what you're sending makes your prompts — and the AI's answers — much sharper.
When to use which of LoL2LLM's five personas, plus the one-line additions that turn a generic answer into a sharp one.
ChatGPT and Claude are excellent at LoL analysis — but they fail in five predictable ways. Here's how to neutralize each with a single prompt addition.