Claude Code and Codex hand everyone the same wall of text. AnswerFit re-renders that answer for the person reading it — the jargon you already know collapses out of the way, and the parts that are new to you bloom into explanations you'll actually get.
You've got a merge storm on your hands. First, abort the in-progress merge, then rebase your feature branch onto main instead of merging — this keeps a linear history and avoids the detached HEAD state you're stuck in.
Enable git rerere so the resolutions get recorded and replayed. Watch out for the TOCTOU race in the deploy hook — it re-reads the ref after the check, so an idempotent retry can double-apply. Force-push with --force-with-lease, never plain --force.
…and 6 more paragraphs, one shape, for everyone who ever asks.
Every reader gets the exact same answer — so it fits almost no one.
An AI assistant has no idea who's on the other side of the screen. So it writes for a phantom “average” developer: enough jargon to sound precise, enough hand-holding to be safe. The result is a wall of text that's simultaneously too much and not enough.
Two branches touched the same lines. Replay each conflict one commit at a time →
$ git config rerere.enabled true
Records how you resolve a conflict, replays it next time.
Fable extracts a concept graph from any AI answer in a single cached call — every piece of jargon becomes a node.
Jev judges each concept against your learner state and returns a calibrated probability that you already understand it.
Known → collapse it. Gap → intervene. Wrong shape → pick a better representation: a DAG, a timeline, a code snippet.
The answer rebuilds around you. One tap of “I already knew this” updates your state and the next answer differs.
One claude-fable-5 call turns an answer into a graph of concepts and dependencies. Results are SQLite-cached, so repeat runs are near-instant and free.
TypeSafe System One scores per-concept mastery with calibrated confidence — ~113 req/s. An FSRS engine keeps state that's always recomputable from an immutable evidence log.
Per concept: already_understands, intervention, or representation. Confidence under 0.5 escalates to Fable. 92.9% agreement when Jev clears the gate.
AnswerFit never guesses from a quiz. It reads the trail you've already left — then lets you correct it in one tap.
Search & watch history. Real Google and YouTube events — searches, visits, watches — matched to concepts and judged by Jev.
Your own transcripts. Your typed messages in Claude Code & Codex, judged as confusion, correct use, or an advanced question.
“Didn't land” signals. A remedial search minutes after an assistant explained something becomes strong negative evidence.
Your feedback. “Already knew this” / “still fuzzy” overrides everything and reshapes the very next answer.
No. It's a personalization overlay. The underlying answer is unchanged — AnswerFit only decides which parts to collapse, which to expand, and what shape to render them in for the specific person reading.
From evidence you already generate: your search & watch history, and your own messages in Claude Code and Codex. Everything is stored locally in SQLite as an immutable evidence log, and your state is always recomputable from it.
Jev is TypeSafe's System One — it returns calibrated probabilities that you understand a concept. Fable (claude-fable-5) extracts the concept graph and writes the fitted explanations. Low-confidence Jev calls escalate to Fable.
Yes. There's a live capture loop: a Claude Code Stop hook feeds every turn back in as evidence, and one tap of feedback updates your state immediately. The next answer is fitted to the you that exists now, not last month.
Yes — it's an open-source Next.js app. Clone the repo, add an ANTHROPIC_API_KEY and JEV_API_KEY, run the backfill, and npm run dev. Setup steps are in the README.