AI as Tool AND Distraction — The 2026 Specific Skill

Use AI coding and writing tools as actual productivity instruments, not as elaborate procrastination. Know in advance which task classes AI accelerates (well-scoped boilerplate, scaffolding, unfamiliar-library orientation) and which it slows (unscoped exploration, novel debugging, mature codebases you know well). Build the personal calibration — from your own logged data — that lets you choose tool-vs-no-tool consciously instead of reflexively. AI tool use is a CI/CD pipeline with quality gates. Some changesets benefit from the pipeline (boilerplate, scaffolding); others should bypass it (small fixes verifiable by inspection). The mistake is treating the pipeline as universal — routing every task through AI because the tool is open in a tab. The skill is the gate, and the gate is a pre-commit hook: it runs before the work starts, not after three hours of toggling. The toggling pattern is the AI-era Slack-check: same compulsion loop, different surface. Compulsive ChatGPT-toggling provides intermittent dopamine that resembles progress without producing progress — and unlike Slack, it produces artifacts (chat transcripts, half-integrated suggestions) that make the procrastination look like work product. The cure is the same as M7's: rate-limit the interactions, pre-commit to the actual task, consult the tool deliberately rather than falling into it. And the calibration illusion is the deepest lesson: the METR subjects finished the tasks and still misestimated the direction of the effect by 39 percentage points. Your introspective sense of whether AI is helping is not data. The log is data. This is M5's lesson — measure, don't introspect — applied to the newest variable in your stack.