quant-dev-prep-skill
A released Claude Agent Skill for quant developer interview prep, built for SWE-background candidates with little or no finance knowledge. It keeps a longitudinal record of your mistakes across four tracks and turns every review session into targeted practice.
SWE-style interview prep doesn't transfer cleanly to quant. LeetCode grinding gets you through the coding round; it does not prepare you for the deep-dive chain a quant interviewer runs after — complexity justification, cache and memory-layout questions, concurrency behavior, and probability puzzles that require you to state both the answer and the reasoning.
And the finance context — order books, fills, market making — is usually taught by osmosis, which fails candidates coming from a pure SWE background.
Four practice tracks: (A) algorithm coding in C++, (B) C++ language internals, (C) systems and computer architecture, (D) probability and brainteasers. Every session writes to persistent logs — mistakes.md, qa-log.md, postmortems.md, idioms.md — so weak concepts get surfaced again on the next review.
Mistakes aren't marked cleared when the user says so — they're marked cleared when the user solves a fresh variant that stresses the same failure mode with a different problem surface. Real-interview mistakes (logged from actual post-mortems) carry double weight in the review queue.
Mock interviews follow the 45–60 min quant round shape: warm-up C++ rapid-fire, main coding problem, deep-dive chain (complexity → cache/memory → concurrency), a systems or probability closer, then debrief with a hire signal (strong hire / hire / lean no / no). Firm-aware — if the user names DRW, IMC, Optiver, Jane Street, HRT, Citadel, or SIG, the mock skews toward that firm's known emphasis.
And an industry-primer track: any finance concept is translated into SWE terms first, jargon second. No assumed knowledge of order books, fills, or basis points.
Shipped as a Claude Agent Skill, MIT licensed. The design choice worth naming: mistake tracking is longitudinal and append-only. Nothing gets deleted, statuses only update. That turns the log into a study surface, not a to-do list.
Tone is deliberately blunt — like a good staff engineer doing code review. If a solution is fine, the skill says "correct, clean, move on" and spends the budget on follow-up questions instead. Praise padding gets in the way of learning.