Interactive, synthetic-but-grounded illustration
How AI reshapes engineer productivity at Anthropic
Built from the article’s reported shifts: Claude usage 28% → 59%, self-reported boost
+20% → +50%, “new work” 27% of Claude-assisted, papercuts
8.6%, autonomy 9.8 → 21.2 tool-call chains.
Before ↔ After
Slide or press play
Before
After
Role → Task expansion flow
Metric: “Successful task outcomes per engineer-month” flowing from role to task type.
Capability bloom
Metric: % of engineers reporting “I can capably ship here” across domains.
Less time, more output
Metric: self-reported time change vs output change by task category (Figure 2 style).
Cross-stack contribution matrix
Metric: % of each role shipping work in each domain (heatmap morph).
Breadth “fireflies”
Metric: domains touched per engineer-month (132 engineers, animated swarm).
Claude task-mix compass
Metric: Claude Code task distribution shifts toward complex work (Figure 4 vibe).
Feedback loop shrink
Metric: end-to-end iteration cycle time (hours) across plan → implement → check → fix → coordinate.
Autonomy dashboard
Metric: task complexity, max consecutive tool calls, and human turns (Feb → Aug 2025).
Time-to-competence distribution
Metric: days to ship a first “unfamiliar-domain” change (synthetic, aligned to interviews).
“A million horses” branching
Metric: parallel hypothesis exploration (branches grow + particles race).
Turn-taking barcode
Metric: human turns shrink while autonomous tool chains lengthen.
Learning curve shift
Metric: “effective proficiency” climbs faster with tighter feedback loops.
Work iceberg
Metric: share of “quality-of-life” work (docs/tests/tools/papercuts) that rises above the priority line.
Papercut backlog burn-down
Metric: open papercuts over 12 weeks; “papercut fixes” ≈ 8.6% of Claude Code tasks.
Work portfolio shift
Metric: output split into “core”, “accelerated core”, and “new otherwise-not-done”.
Quality-of-life flywheel
Metric: papercut fixes reduce friction → frees time → fixes more papercuts.
Toil budget waffle
Metric: “activation energy” drops, reallocating hours into neglected work.
Backlog age raincloud
Metric: completed “neglected” items skew younger (stale tail shrinks).