Out of The Crisis: Convergent Evolution and AI-First Design
How smart systems thinking always takes us to the same logical place
Everyone lost their minds over Boris Cherny’s Claude Code workflow. With good reason: it’s amazing.
The creator of Claude Code posts a thread on X about how he runs 5 terminal sessions in parallel, uses Plan Mode, keeps a shared CLAUDE.md, and the entire developer internet treats it like scripture. VentureBeat covered it. Slashdot covered it. Medium exploded with “13 Practical Moves” breakdowns.
Boris Cherny's original workflow was shared on X in January 2026. Coverage from VentureBeat, InfoQ, and paddo.dev is worth reading for the full engineering perspective.
What surprised me? I had created a similar workflow without even knowing, only from the service/product designer lens!
Not the same way. Not from the same direction. But the same destination. That convergence is the interesting part to me. See, Boris approaches this as an engineering optimization problem. Maximize throughput, minimize correction cycles, compound knowledge in code. He is a Staff Engineer at Anthropic. His instinct is to make the machine go faster.
I approach this as a service design system. Specialized roles, bounded responsibilities, voice as identity, orchestration as craft. I am a designer who spent years learning how humans and systems interact. My instinct is to make the collaboration legible.
We arrived at the same place from opposite ends of the map. That is convergent evolution, and it tells you something important about where all of this is going.
But Before we start…
Put W Edwards Deming’s principals from his seminal book on organization design, Out of the Crisis, in your mind as you read. You’ll understand why at the end:
Create constancy of purpose for improvement.
Adopt the new philosophy. Accept that poor quality is unacceptable.
Stop relying on inspection to achieve quality.
End the practice of awarding business on price alone.
Improve the system continuously.
Institute training on the job.
Institute leadership, not supervision.
Drive out fear.
Break down barriers between departments.
Eliminate slogans and exhortations.
Eliminate numerical quotas and targets.
Remove barriers to pride in workmanship.
Institute education and self-improvement.
Put everyone to work on transformation.
The Converging Paths
Parallelism
Boris runs 5 numbered terminal tabs with generic sessions. Clean, fast, interchangeable.
I run named agents with domains: Siddig owns backend, Heavy owns frontend, Decker owns the prose IDE. Each one has a personality, a voice, and a bounded responsibility. Think specialist teams versus generalist pools. In service design terms, this is the difference between “anyone can work the counter” and “this person owns this touchpoint.”
Isolation
Boris uses git worktrees or separate checkouts to keep sessions from stepping on each other.
I use separate Warp terminal panes, each with its own CLAUDE.md. Bounded contexts, like in a service blueprint: each agent knows its territory and stays in its lane.
Memory
Boris maintains a single shared CLAUDE.md, about 2,500 tokens, checked into git. The whole team contributes to it.
I maintain a per-agent CLAUDE.md with voice, domain expertise, crew relationships, and working style. Julian’s CLAUDE.md alone is longer than Boris’s entire shared file. This is the difference between a system map (one view of the whole) and journey maps (one view per role).
Planning
Boris uses Plan Mode: iterate on a plan with Claude, then switch to auto-accept and let it execute.
I use Julian (the orchestrator) to plan, then hand mission briefs to the executing agent, then Qin (the auditor) reviews. Discovery, definition, delivery. Three roles, three phases, three gates.
Review
Boris has a second Claude review the plan “as a staff engineer.” Adversarial by design.
I have Qin, the Code Inquisitor, gate all merges with a formal audit protocol. Classification levels, commendations, findings by severity. QA is not an afterthought. It is a first-class role with its own voice and its own standards.
Error Recovery
Boris’s rule: when Claude makes a mistake, update CLAUDE.md so it does not happen again. Simple and effective.
My version: the agent writes the correction into its own CLAUDE.md, and the correction gets archived to Notion for cross-agent visibility. Blameless postmortems that become updated playbooks. The learning persists beyond the session.
Prompting Style
Boris’s signature prompt: “Knowing everything you know now, scrap this and implement the elegant solution.” Direct. Engineering-brained. Gets results.
I dictate mission briefs with character voice and emotional context. The agents do not just know what to build; they know why it matters and who they are while building it. Co-design sessions versus requirements docs.
Automation
Boris checks slash commands into
.claude/commands/and runs/commit-push-prdozens of times a day. Ritual automation, codified and versioned.
I formulate prompts through Julian and hand them off manually. Less automated, more orchestrated. The tradeoff: I lose speed, I gain legibility. (This is the one where Boris is clearly ahead. I need to build my command library.)
Voice Input
Boris uses macOS dictation (fn key twice). Quick, built-in, 150 words per minute.
I use VoiceInk for transcription and dictate prompts as standard practice. Dictation is not a convenience for me; it is a design methodology. Speak first, structure later. Participatory design applied to AI prompting.
Bug Fixing
Boris pastes a Slack thread and says “fix.” Minimum viable context.
I describe the symptom to the right agent, with system context. The agent who owns that domain gets the brief. Triage by domain expertise, not by whoever is closest.
Context Hygiene
Boris uses subagents to offload work and keep context windows clean.
I use separate chat threads per concern and agent-specific conversations. Channel management, like in any service operations center. Keep the signals separated so the noise does not compound.
Verification
Boris’s golden rule: “Give Claude a way to verify its work.” Browser testing, test suites. He says this alone 2-3x quality.
I run servers, click through flows manually, and have Qin do post-audits. Usability testing: the proof is in the doing, not the telling.
Identity
Boris runs anonymous sessions. Numbered tabs. The AI is a tool.
I run characters with pronouns, voices, relationships, and backstories. The AI is a collaborator with a role to inhabit. Personas are not just for users. (This is the single biggest philosophical difference between the two approaches, and I think it matters more than anyone realizes yet.)
Knowledge System
Boris keeps CLAUDE.md plus slash commands checked into git. Clean, portable, version-controlled.
I keep CLAUDE.md plus Notion archival plus Claude Projects plus migration history documents. Institutional knowledge as living documentation, not just code comments. More overhead, more resilience.
Model Choice
Boris uses Opus 4.5 (and now 4.6!) for everything. Pay the compute tax, skip the correction tax. Smart.
I use Opus 4.6 for Julian and strategic work. The agents in Claude Code run on whatever model Claude Code serves. Right tool for the right moment. (Boris is probably right that Opus everywhere is the better bet. But my workflow does not give me that dial to turn for the agents.)
The Throughline
Boris and I are solving the same problem: how do you manage AI agents at the scale where a single person operates like a team?
He solved it like an engineer. Optimize the pipeline. Minimize friction. Compound learnings in code. Make the machine go faster.
I solved it like a designer. Define the roles. Bound the responsibilities. Give the agents identity. Make the collaboration visible.
The real insight is that managing AI agents is management. Full stop. Delegation. Specialization. Review gates. Knowledge transfer. Error recovery. These are not new problems. They are the oldest problems in organizational design, wearing a new hat.
The question is not “how do I use Claude Code better.” Rather, it is “how do I run a team.” If you have ever managed people, you already know more about this than you think. And if you have never managed people, well, now you are. Surprise. You are designing an organizational playbook. You’d be better off reading W. Edwards Deming “Out of the Crisis” than books on AI at this stage.
Boris just showed the engineering playbook. I have been running the design playbook. The fact that we converged tells you the fundamentals are real.
Now go build your crew. Need help? Try my Investiture framework to get started!