Over five months (Feb–Jun 2026) you asked 700 questions across Claude Code and Codex. It started with short "how do I…" lookups — roughly 34 characters each — and by late spring the average question grew to ~272 characters: multi-step specs with code blocks and explicit goals. The vocabulary moved with it, from setup basics (docker, git, pip) to system concerns (latency, cache, throughput). And the way you work with the AI matured — 300 trade-off/critique prompts and 270 verification checks: you started directing and pressure-testing answers, not just dictating tasks. (This is a sample report built from synthetic data.)
Reading the current patterns to find the next focus.
Code blocks show up in ~39% of questions, but almost all of them are in the later months — the early asks were context-light.
Paste the error and the relevant file at the start of a question to cut the back-and-forth before the AI has enough to work with.
Verification asks reached 270, but counter-questions ('why not…', 'rationale') stayed low at 90 — you check answers more than you challenge them.
After a proposal, ask 'why not the simpler option?' to force the AI to defend the design before you commit to it.
Repeated, re-explained tasks — build them once as a skill and stop re-asking.
You ask for the same trade-off comparison over and over — re-describing the options, the goals, and the constraints from scratch each time instead of reusing a fixed frame.
/skill-creator: build an 'architecture trade-off review' skill. Input: two candidate options + the goal axes (latency / cost / ops load). Output: a side-by-side table on latency, cost, operational load and risk, then a recommendation and the verification questions to run. Representative asks: "kubernetes vs a plain vm for the billing service — what are the trade-offs?", "why not just scale the workers horizontally?".
Latency and throughput investigations repeat with the same shape — current numbers, a budget, then 'walk me through the options' — and you rebuild that context every time.
/skill-creator: build a 'latency triage' skill. Input: current p50/p99 + target + a one-line topology. Output: where the time likely goes, two architectures with trade-offs, and a check against the availability target. Representative ask: "p99 ~800ms, goal under 200ms, keep the cache warm after deploys — walk me through the trade-offs and verify the plan against the goals."
Bar height = monthly question count · number = count for that month
avg questions / session 11.29 · one-shot (single-question sessions) 2% · sessions 62
Question genres — Build 31% · Debug 30% · Understand 17% · Other 17% · Improve 4%
The phases where your questions shifted.
Short, practical questions — installing tools, fixing startup errors, basic git. Average ~34 characters. You were learning the terrain.
Questions doubled in length and turned comparative: kubernetes vs a VM, redis vs memcached, and the first 'why is this slower?'. You started asking for trade-offs.
Multi-step specs with code blocks, latency budgets, and 99.9% availability targets. You set the goals, asked for trade-offs, and made the AI verify its own plan.
Cross-sections of your working style revealed by your questions.
Your questions migrated from 'how do I set this up' to 'how should this system behave under load'. The early months are full of tooling basics; by May the top terms are almost entirely about runtime behavior.
The shift from procedure to principle is the clearest signal in your data: questions got ~8× longer as you moved from 'how do I' to 'why is this slower' to full design trade-offs.
You went from taking dictation to running the room — asking for trade-offs, demanding verification, and delegating the final call while keeping the reasoning visible.
Your prompts became self-contained: code blocks for context and numbered steps that hand the AI a plan instead of a fragment — fewer round-trips per intent.
Some topics you asked about intensely, then stopped — a sign you internalized them and moved on.
Three seasons of work surface on their own: a setup season, a comparison season, and a scale season.
Click "Save as image" on each card to download it as PNG.