Games in view
Product Extension Concept
Track every Chess.com game and turn the season into a coaching story.
A season-level layer for RookLook: sync a player's archive, evaluate every game with engine-first and deterministic signals, then use rolling LLM windows to explain what is actually changing.
Demo Tracker
Preview a player-shaped coaching dashboard
This prototype uses the local coaching corpus in this repo. Type a handle that exists in the corpus, or fall back to the default demo player if we do not have enough games for a real rolling window.
Rolling window
Nightly archive sync
Stockfish plus deterministic enrichment
Rolling LLM summaries, not one-shot verdicts
Trend tracking across openings and habits
Loading local coaching corpus...
Record
-
Castled by move 10
63%
Wins from +1.5
54%
Tracking @spelunkid
LoadingCurrent coaching arc
Loading corpus-backed coaching view
Next training block
What to work on next
The page should move from analysis to action. Every narrative window ends with a concrete practice block, not just a diagnosis.
Narrative ledger
How the story changes across windows
Instead of asking one model to explain a whole season at once, summarize stable windows and compare them for drift, growth, and repeated failure modes.
Repeated themes
Habits backed by engine evidence
These themes come from computed features like king shield, center control, rook activity, and development tempo, not from the model inventing board claims.
Repertoire view
Where the trends actually live
The extension should separate broad habits from opening-specific issues so the player can tell whether a problem belongs to the repertoire or to the player.
Core guardrail
Never let the model invent chess
The same principle already exists elsewhere in this repo: the engine and deterministic annotations do the chess reasoning, and the LLM turns structured evidence into a readable coaching narrative.
"The model never sees raw geometry. It sees eval swings, king safety, center control, open-file rook usage, pawn structure, and trend aggregates."
That makes the page safer and more useful. It also creates a clean contract for windowed summaries, longitudinal comparisons, and eventual coach-facing reports.
Suggested pipeline
How the real system would work
This is the product shape behind the page: a stable ingest and analysis layer feeding a narrative layer that only summarizes computed signals.
Archive ingest
Pull monthly Chess.com archives for a player, dedupe by game id, and group by time control and color.
Non-LLM evaluation
Run Stockfish plus deterministic enrichment for features like center control, king shield, open files, piece mobility, and pawn structure.
Rolling windows
Build last-5, last-12, and last-30 style aggregates so a single blunder does not rewrite the player's story.
Narrative synthesis
Feed the model only structured summaries, recurring motifs, and opening buckets so it can explain trends without hallucinating positions.
Coach output
Surface the current arc, stable leaks, what improved, and a short practice block that changes as the window changes.
Longitudinal memory
Save each window summary so the next pass can compare against prior windows and speak in trends instead of isolated games.