Understand it.
Then explain it.
First-principles guides to the AI concepts that come up in serious technical interviews—built with diagrams, code, and numbers you can reason through.
The study library
12 topics
RAG
How a model finds useful evidence before it answers—and where retrieval systems quietly fail.
Attention
The query–key–value mechanism, worked by hand before we touch a single line of PyTorch.
RLHF
How human preferences become a training signal—from pairwise labels to a safer policy.
KV Cache
Why autoregressive generation would be painfully wasteful without cached keys and values.
Evals
How to measure whether an LLM system is actually getting better—without gaming your test.
FlashAttention
Why attention is limited by memory movement, not math — and how tiling plus an online softmax fixes it.
Rate Limiters
How APIs say "slow down" without falling over — token buckets, sliding windows, and the distributed race.
Continuous Batching
How inference servers stop wasting GPU on padding — iteration-level scheduling and paged memory.
Model Degradation
How production models quietly rot — drift detection, golden evals, and the monitors that catch it.
URL Shortener
Shrink URLs without breaking the internet — base62 counters, the key service, and 91 terabytes of storage math.
Distributed Cache
The senior-to-staff question — consistent hashing rings, 3x replication, eviction, and taming the thundering herd.
YouTube
Two systems wearing one logo — async uploads, transcode farms, and why hot videos live on the edge.
Twitter Timeline
Push vs pull fan-out — precompute timelines for the many, merge celebrities in at read time.
Load Balancer
L4 vs L7, five algorithms, health checks, and the failure math — the traffic cop every backend interview expects.
Training-Data Pipeline
Crawl the web, keep the best tenth, kill duplicates, scrub PII — how LLMs actually get their data.
Claude guides
from @theclaudecraft
Anatomy of a .claude/ folder: 7 things inside
Every Claude Code project hides a .claude/ folder that quietly decides how Claude behaves: shared settings, private overrides, skills, subagents, slash commands, memory, and hooks — with the exact file templates.
→7 hidden Claude Code settings that change everything
Buried in the settings files: the permission mode new sessions start in, hooks that guard your secrets, a custom status line, your default model, output styles, and auto-compact — with the exact JSON for each.
→The code-modernization plugin, explained
Anthropic's official legacy-modernization plugin: an enforced assess → approve → build → prove sequence, five specialist agents, three build methods, and a byte-level proof that the new code behaves like the old.
→The security-guidance plugin, explained
Anthropic's official security plugin: instant pattern warnings on every edit, an Opus 4.7 diff review each turn, and an agentic data-flow review on commit — with honest limits.
→Skills vs subagents vs hooks
The three words everyone mixes up: reusable playbooks, fresh-context specialists, automatic tripwires — and which one you actually need.
→5 official Claude plugins you didn't know existed
Anthropic's own marketplace plugins: deep security scans, codebase setup plans, CLAUDE.md care, the agent SDK kit, and real C++ code intelligence.
→Spotify's shunt plugin, explained
How Spotify cut Claude Code token usage by a reported 90%: bulk reads and boilerplate rerouted to a cheap worker model — with honest limits.
→Opus 5.5 in plain English
Anthropic's new flagship: 20% cheaper tokens, 30%+ faster output, a 1M-token context window — and the benchmarks behind it.
→The harness that won Anthropic's hackathon
How 286 skills and 68 subagents work as a pipeline — and the stealable patterns (review loops + memory vault) behind 260K GitHub stars.
→Browse the full Claude library
Every Claude guide in one place — new features, hidden GitHub tools, and agent harness patterns, explained simply.
→Learn the mechanism,
not the buzzword.
Every guide starts with the smallest useful mental model, makes the arithmetic visible, then climbs toward production trade-offs. Interview prompts are framed as practice—not leaked question banks—so you learn to reason under follow-up, not memorize a script.