A Year With Claude Code: Output Up, Attention Span Not

Key takeaways

1. Date / source. 5 May 2026, LinkedIn Pulse. Author: Alexandre Frizzo. URL: https://www.linkedin.com/pulse/year-claude-code-my-output-doubled-attention-span-didnt-frizzo-atmoe/
2. Format. one-year retrospective — a year of daily use of Claude Code, an honest assessment.
3. Title-aphorism. "My output doubled. My attention span didn't." — a perfect symmetry that condenses the diagnosis.
4. Central thesis. "The new bottleneck is supervision. Reading what an agent produced, deciding if it's right, integrating it, catching the subtle things it got wrong." ### The job has changed shape
5. Before. "write hard parts, review easy parts" (the developer writes the difficult logic and reviews the details).
6. After. "deciding about code" across multiple simultaneous codebases — the developer has become a supervisor of agents.
7. The bottleneck has shifted.
8. Old bottleneck. protecting deep work state (entering flow, deep concentration).
9. New bottleneck. high-speed supervision (reading agent output, deciding correctness, integrating, catching subtleties).
10. "The traditional focus on protecting deep work state became irrelevant once code generation became fast."
11. "Quality control at scale now requires entirely different defenses." ### Acknowledged gains
12. 3-5× productivity. on a typical day. This figure is consistent with Wescale X3-X4 (2026-05-03) and reflects the experience of a practitioner committed for 1 year. Not to be confused with the elite tail, which goes much higher: Cherny (2026-05) states "a few dozen PRs/day, 150 PRs in a single day record" (clearly >10× the usual baseline), Karpathy (2026-04-29) specifies "10× is not the speed up — people who are very good at this peak a lot more than 10×", Curran/Intercom (2026-04-16) measures that the top 5% of performers generate 6× the median PR throughput. Correct reading: uneven distribution, 3-5× for the committed median, 10×+ for the elite tail. The myth to debunk is not 10× itself but "automatic 10× for everyone with a simple prompt".
13. Previously infeasible projects now achievable. yak-shaving, boilerplate, marginal automations — the long-tail becomes addressable (consistent with Wescale 2026-05-03 "historical needs that remained too costly can finally be addressed").
14. Near-zero cost of experimentation. — the mental barrier to trying things is lifted. ### Acknowledged losses
15. "Writing muscle". atrophied — "manual coding feels effortful now". A bodily metaphor that makes degradation through disuse tangible.
16. Rare deep flow state. — "constant context-switching" between multiple supervisions. The cognitive flow state protected by 30 years of Deep Work literature (Cal Newport) becomes inaccessible.
17. Diminished ownership satisfaction. — "the code is good, but isn't quite mine." The psychological sense of creation is eroded. ### Statistics cited (supporting evidence) | Measure | Value | Source | |--------|--------|--------| | Median effective coding time out of an 8h day | 3-4 hours | Developer research | | Context recovery per interruption | 23 minutes | Classic Gloria Mark study (UC Irvine) | | Entering flow state | 15-25 minutes | Flow research | | Rebuilding mental context | 15-30 minutes | Cognition research | | Productivity multiplier in flow | 500% | McKinsey senior executives study | ### Unresolved tensions (the questions he does not settle) 1. FOMO Problem — knowing that Claude can work overnight creates psychological pressure. "Every hour I'm not at the keyboard is an hour an agent could be earning for me." Weekend/evening work becomes tempting even when projects are finished. An inversion of the deep-work logic: it is no longer the absence of productivity that weighs, it is the absence of active supervision. 2. Review Quality at 3-5× Volume — reviewing at 3-5× the volume risks skimming. The industry's historical quality practices assumed a human pace. At this throughput level, subtle bugs can slip through review. 3. Skill Atrophy — Frizzo poses the open question: "does the writing muscle still matter, or is it becoming commoditized?" He does not settle it — it is a political point. ### Ethical-epistemic position
18. Simultaneous refusal of the two available narratives.
19. "AI is bad" — Frizzo explicitly rejects it.
20. Uncritical enthusiasm — he rejects this too.
21. Third-position stance. real gains + real costs, unresolved tensions rather than conclusions. ### Connection to the watch corpus
22. A welcome counterweight to Cherny's *"coding is solved". (2026-05, fiche cherny-sequoia-coding-is-solved-loops-printing-press-2026-05.md). Cherny declares 100% of code generated, "loops are the future"*, an iOS setup with 5-10 sessions / a few hundred agents. Frizzo, with the same intensive daily use over a year, concludes that the cognitive costs are real. The 2026 corpus needs both voices to stay honest.
23. Numerical convergence — uneven distribution.
24. Committed median 3-5×. (the most widely documented experience):
25. Frizzo. 3-5× productivity multiplier over 1 year of daily use.
26. Wescale Usine Logicielle Augmentée. (2026-05-03): "realistic X3-X4".
27. Curran/Intercom. (2026-04-16): 3× R&D productivity in 16 months (entire organization, 500 R&D).
28. DORA Report 2025. (2025-09-23) and Stanford Denisov-Blanch (2025-11-23): quantify AI ROI.
29. Elite tail 10×+. (practitioners who have entirely reorganized their work):
30. Cherny. (2026-05): "a few dozen PRs/day, 150 PRs in a single day record" (clearly >10× a dev's usual baseline).
31. Curran/Intercom top 5%. 6× the median PR throughput within an organization already averaging 3× — i.e. 18× the pre-AI baseline for this subset.
32. Karpathy. (2026-04-29): "10× is not the speed up — people who are very good at this peak a lot more than 10×".
33. → Correct reading: uneven distribution, 3-5× for the committed median, 10×+ for the elite tail. The myth to debunk is "automatic X10 for everyone", not X10 itself. Frizzo and Cherny do not contradict each other — they measure two different points of the distribution.
34. Convergent cognitive tensions.
35. BCG/HBR Brain Fry. (Bedard et al., 2026-03-05): 14% AI brain fry, peak 3 tools, +39% major errors, +39% intent to leave. Frizzo provides the individual testimony that the BCG study measures at the population level.
36. Anthropic January 2026. study (cited by Sun NYT 2026-04-30): "junior engineers using AI agents understood their work less when quizzed afterward." Frizzo confirms individually: writing muscle atrophy.
37. Karpathy *"outsource thinking but not understanding". (2026-04-29): Frizzo experiences the tension "code is good, but isn't quite mine" which is exactly the loss of understanding while thinking* has been outsourced.
38. Beck *Starving Genies. * (2026-04-03): voluntary scarcity vs FOMO from 24/7 agents. Frizzo identifies FOMO as an unresolved tension.
39. Soto *Developer Taste. (2026-04): taste as the last skill. Frizzo poses review quality at scale* as a challenge to taste at high throughput.
40. Mornati *What is a Developer When We Use Coding Agents?. * (2026-03-14): 1-day experience, the 70% problem. Frizzo confirms after 1 year: the nature of the job has structurally changed.
41. Klaassen *Stop Coding and Start Planning. * (2025-11-06): compounding engineering, three fidelities. Frizzo lives this transition (planning > coding) on a daily basis.
42. Convergence on "supervision is the new bottleneck".
43. Wescale. (2026-05-03): Strategic Judge + Agent Manager — the human becomes a supervisor.
44. Karpathy. (2026-04-29): agents as interns with recall but no taste — the human in charge of spec, judgment, oversight.
45. Andreessen. (2026-02): orchestrating 10 bots.
46. Habert PROJ-AI. (2026-05-05): "agent directives + Decision Records + five validation dimensions".
47. Frizzo. experiences the supervision bottleneck day to day and names the costs.
48. FR vs Anglo-Saxon position.
49. Frizzo writes in English on LinkedIn — not specifically French despite the name.
50. His epistemic honesty aligns more with French caution (Wescale, Habert) than with American optimism (Cherny, Curran).
51. To be used in FR presentations as a practitioner testimony complementing Wescale/consulting-firm figures. ### Limitations to flag
52. Author identity poorly documented. Alexandre Frizzo — no visible institutional credentials, no cited industrial practice. Authority built through the text's coherence, not through external credentials.
53. Sample n=1. a year of personal use, not a study. But methodological honesty compensates — he cites external studies (Gloria Mark, McKinsey).
54. No verifiable internal figures. on the 3-5× productivity — it is a subjective estimate, not an output measurement (vs Curran/Intercom, which measures PRs).
55. No action roadmap. — the text is diagnostic, not prescriptive. Those seeking recommendations should complement it with Wescale, Habert, Klaassen. ### To be used for
56. Executive committee presentations. balanced, anti-hype practitioner testimony that qualifies X10 promises without demobilizing.
57. Manager awareness. of the hidden cognitive costs of AI transformation (FOMO, deskilling, ownership erosion).
58. Ethical debate. on the transformation of the developer's role into a supervisor.
59. Sourced figures. 3-5× realistic, 23 min Mark, 500% McKinsey flow — figures to integrate into watch fiches, strategic presentations, internal training.