@bcherny

Engineering Workflow & AI Agent Strategy

Cherny's through-line is that the old engineer habit of automating busywork (vim macros, lint rules, e2e tests) has become even more leverage now that an 'army of agents' benefits from the same speed-ups, and that the domain knowledge which used to live in people's heads should be encoded as infrastructure so both agents and non-expert humans can contribute on day one [doc 1]. His sharpest formulation: 'Every team should be writing the CLAUDE.md's, REVIEW.md's, skills, and docs that enable agents to productively work in their codebase with zero additional context from the prompter' [doc 1] — calling it a 'natural extension of the stuff engineers have always done' [doc 1]. He concedes Claude itself should get better at proactively managing long-term memory and says it is 'improving quickly' [doc 1]. In his most-discussed post — 65 replies, ~70k views — he introduces a 'Steps of AI Adoption' framework, observing that 'one person is 10x'ing their output with Claude but the rest of the org hasn't caught up' and that moving up each step requires breaking the next bottlenecks and building the next guardrails, not just spending more tokens [doc 2]. Practically that means giving Claude ways to verify its own work end-to-end, enabling auto-mode permissions, defaulting on automated code and security review, and using multi-agent interfaces (CLI Agent view, Desktop, iOS/Android, Tag); higher levels add /loop, /batch, dynamic workflows, and worktree isolation for subagents [doc 2]. On measurement he is explicit that usage dashboards measure 'activity, not return' and that the better question is 'would you have spent engineering effort on this anyway? … how much and what would it have cost in manual eng-hours? That's your return' [doc 2]. The payoff arrives when 'fixing and maintaining happens in the background and your teams can focus on building … things that weren't even in range before'; he places Anthropic at step 3 pushing toward 4 and himself personally at level 4 [doc 2].

Domain knowledge as infrastructure: Argues encoding team knowledge in CLAUDE.md, REVIEW.md, skills, docs, and memories is the modern equivalent of writing lint rules and e2e tests — it multiplies agent output, automates entire classes of busywork forever rather than one-off fixes, and lets non-engineers and newcomers contribute on day one with zero extra prompting context [doc 1].

Steps of AI Adoption framework: Maps a four-step adoption ladder, stresses that tokens alone don't unblock the next step (bottlenecks + guardrails do), lists the specific features that enable each level (auto-mode, automated review, /loop, /batch, worktree isolation), and reframes ROI as 'would we have spent eng-hours on this anyway?' rather than usage metrics [doc 2].