Measure — and train — how well you actually work with AI, from your real Claude Code sessions. No quizzes, no self-report: your behavior is the input.
Most people can't feel whether AI is making them sharper or lazier. In a controlled RCT, experienced developers believed AI made them ~20% faster — measured, they were 19% slower, and couldn't tell (METR, 2025). Separately, a ~1,000-student RCT found unrestricted AI made students 17% worse on unassisted exams, while the same tool with guardrails caused no harm (PNAS, 2025). The difference isn't the tool — it's a set of observable behaviors. You can't self-assess them, so we measure them, then help you train them.
A Claude Code plugin records lightweight events from your real sessions to a local file (~/.ai-fluency/events.jsonl). It is fail-open and fast — it can never slow or break a session — and it only scores your own interactive sessions, never agent or scripted runs.
After each turn, an on-device engine reads the transcript and scores 7 behaviors 0–100 using deterministic heuristics (plus an optional local LLM judge for the judgment-heavy ones). Everything is computed on your machine.
You get a one-line “what to do better” in the terminal statusline and live coach, plus your single weakest habit to focus on. Optional in-flow nudges are observe-only until you opt in.
Aggregate scores (never your prompt text) sync to your dashboard and a shareable profile — a radar of your behaviors, your trend, and your Fluency Rating.
Each is scored from your real sessions and grounded in the research. Weights reflect the evidence — verification and diagnosis matter most.
Fit-to-task, not a ladder. A behavior that doesn't apply to a turn is marked not-applicable — never penalized. A calibrated one-line prompt is not "weak." And nothing scores speed, volume, or acceptance-rate — only judgment.
Your headline number (0–100, with a band from Emerging to Expert) is the mean weighted quality of your recent turns. It is deliberately volume-independent: grinding more sessions can't inflate it — only working better can. It stays provisional until you have 15 scored turns, so a number you show a recruiter reflects a real sample. A separate Level/XP tracks activity — that's the game layer, not the credential.
This is a telemetry tool from a solo publisher, so don't take it on faith — the whole path is open source (MIT) and built to be audited.
text and file paths are stripped twice: client-side before sending, and again server-side on receipt.sync.py --dry-run.Today your profile is a self-instrumented behavior profile, not an audited credential — it's computed from your own machine's telemetry. A formal validation study (tying these behaviors to real learning outcomes) is in progress. We don't cite the "70% get dumber" figure some people repeat — it isn't sourced; the honest number is that roughly 40–50% of real AI use is passive delegation. We'd rather under-claim and let the measurement speak.
MIT licensed · local-first · public source repo at launch