Detailed Analysis
Boris Cherny, the creator of Claude Code at Anthropic, has systematized his personal development workflow into a publicly shareable skill file — a structured SKILL.md document hosted at howborisusesclaudecode.com — that encapsulates 72 workflow tips across 60 sections, designed to function as an in-session coaching layer within Claude Code itself. The document, compiled by a community contributor and versioned at 5.0.0 as of April 2026, operates in two modes: a proactive mode that surfaces contextually relevant tips during active work, and a browse mode triggered via the `/boris` slash command. The skill's architecture reflects a broader philosophy in the Claude Code ecosystem — that reusable, composable knowledge artifacts, rather than rigid scripts, are the scalable unit of developer productivity.
The workflow practices Cherny documents are notable for their scale and parallelism. He runs five simultaneous Claude Code sessions across local terminal tabs, each backed by a separate git checkout, while maintaining five to ten additional sessions on claude.ai/code — a throughput model that enables him to ship 20 to 30 pull requests daily. Central to this approach is plan mode, which delays auto-execution while a plan is iterated and refined, effectively front-loading cognitive investment before implementation begins. The shared CLAUDE.md file — approximately 2,500 tokens, committed to version control — functions as a living institutional memory for the team, updated each time Claude produces an error, converting one-time mistakes into permanent context that compounds over time. This pattern of treating documentation as an error-correction mechanism rather than a static reference represents a meaningful shift in how teams might manage AI-assisted development at scale.
The tooling layer Cherny describes — subagents, hooks, slash commands, and MCP integrations — reflects a mature philosophy of decomposition. Repetitive high-frequency operations like committing, pushing, and opening pull requests are encapsulated in `/commit-push-pr`, while more complex post-task operations are delegated to specialized subagents such as `code-simplifier`, `verify-app`, and `build-validator`. Hooks, particularly PostToolUse hooks for code formatting, automate polish that would otherwise generate CI failures. The skill file's context map — a decision table mapping user behaviors to relevant tips — is itself an example of the metadata-rich, machine-readable design patterns increasingly common in advanced AI workflow tooling, where the agent's behavior is shaped not just by prompts but by structured, queryable knowledge artifacts.
The broader significance of this document lies in what it reveals about Anthropic's internal dogfooding culture and the emergent practice of workflow-as-artifact. Cherny is not merely sharing productivity hacks; he is demonstrating a paradigm in which the developer's accumulated knowledge about how to work with an AI system becomes itself a distributable, versionable, installable module. The one-liner `curl` installation pattern for skills — borrowed from the Unix tradition of portable tooling — lowers the barrier to adopting expert workflows, effectively allowing any developer to bootstrap from Cherny's accumulated experience. This signals a potential future where high-performing AI-augmented workflows are shared and remixed as openly as open-source libraries, with version numbers, changelogs, and community contributors.
The document's self-updating mechanism — instructing Claude to fetch a remote version endpoint and alert users to newer releases — illustrates a recursive quality to the project: the skill teaches Claude how to maintain itself. As AI-assisted development matures, the meta-layer of managing AI behavior, context, and tooling is becoming as technically demanding as the underlying software work itself. Cherny's public documentation of his own practices suggests that transparency about how leading practitioners actually use these tools may become a competitive and cultural differentiator, both for individual developers seeking to maximize throughput and for AI companies seeking to demonstrate the ceiling of what their systems can enable.
Read original article →