Claude Code’s creator, Boris Cherny, has spent 2026 turning an internal Anthropic workflow into a public operating model: teach the agent from its mistakes, verify by running the product, prefer auto mode when work is scoped, promote recurring jobs into routines and loops, and supervise parallel agents from a single control surface. In early November, that playbook lit up developer X again — six-step blueprints, overnight subagent recipes, and arguments about whether to delete your CLAUDE.md every six months. The tutorials vary; the underlying architecture does not.
Quick Take
- What: Developers are packaging Cherny-style Claude Code setups into shareable tutorials: memory files, specialist roles, verification loops and parallel agents.
- Core pattern: Correction → durable instruction; “done” means the agent can run and test the change; recurring work becomes a routine or /loop.
- Official footing: Anthropic documents CLAUDE.md, skills, routines, /loop, /goal, worktrees, subagents and agent view — the viral posts mostly remix those primitives.
- Why it matters: Coding-agent value is shifting from clever prompts to team systems that remember, verify and watch a queue.
What happened
Cherny and product lead Cat Wu have described Claude Code less as a chatbot and more as an operating layer: mistake-to-memory loops, verification that opens the real app, auto mode to cut approval spam, routines that babysit PRs and stale bugs, and agent view plus Remote Control for fleet supervision. Summaries such as The Neuron’s breakdown stitch those habits into a concrete setup list. On November 3–5, builders recirculated that list as paste-ready tutorials — including six-step “install → memory → roles → verify → loop → review” cards — after Cherny-adjacent talks and clips spread on X.
Background: from prompt tips to agent OS
Claude Code can plan, edit files and run commands until a task finishes; Anthropic markets that loop as the product. The 2026 shift is organizational. Shared, git-checked-in CLAUDE.md files turn corrections into compounding context. Skills encode procedures without bloating every session. Routines and scheduled /loop jobs watch issue queues. Worktrees isolate parallel agents. That stack is why “how Boris uses Claude Code” posts outperform one-off prompt packs: they describe a system, not a magic sentence.
How the six-step blueprints usually look
Community cards converging this week typically map to: (1) install and allowlist tools; (2) create a lean shared CLAUDE.md (and optional .claude/rules/); (3) save repeated fixes as memory or skills; (4) define specialist subagents (builder vs verifier); (5) require a real verification path — tests plus running the UI; (6) promote recurring work to routines/loops and supervise in agent view. Cherny’s sharper advice, amplified again this week, includes deleting stale CLAUDE.md/skills periodically so rules written for weaker models do not hold back newer ones — a reset, not a rejection of memory.
| Blueprint layer | Typical Cherny-style practice |
|---|---|
| Memory | Shared CLAUDE.md + skills; write lessons after repeated mistakes |
| Verification | Agent must run/test the change (not only lint) |
| Autonomy | Auto mode when scoped; humans review high-risk moments |
| Recurrence | Routines, /loop, /goal instead of re-prompting |
| Parallelism | Worktrees + agent view; maker vs verifier subagents |
| Hygiene | Periodically delete stale rules written for older models |
Pros and cons
For users
- Pro: Clearer agent setups mean fewer abandoned coding sessions and more reproducible results from the same repo.
- Con: Copy-pasting a six-step card without a verification path produces confident slop at higher speed.
For developers
- Pro: Memory files, worktrees and maker/checker splits are portable patterns across Claude Code and similar agents.
- Con: Parallel agents raise review bandwidth and token cost; the bottleneck becomes human attention, not typing speed.
For companies
- Pro: Routines on CI, stale tickets and PR babysitting can cut toil if scoped with audit logs and approval gates.
- Con: Cloud routines that act as the user need tight repository, network and connector scopes — tutorials rarely emphasize threat models.
What X is saying
Builder posts treated Cherny’s habits as a checklist rather than a vibe.
This is f*cking insane. Boris Cherny, the creator of Claude Code: “Overnight subagents do deeper work” A builder makes the change. A verifier checks the real result. Memory keeps the receipt.
— sharfovii (@sharfovii) November 5, 2026 · View on X
AI notes (@BoboLinNotes)
— Bobo Lin Boris Cherny, the creator of Claude Code, suggests deleting your CLAUDE.md, skills and hooks every six months to see what the latest model can do without them.|November 3, 2026 · View on X
Every time you retype the same instructions to Claude, you are doing a job that a file should be doing. This six-step build removes that repetition one layer at a time.
— Mr. Claude (@claudeskills101) November 5, 2026 · View on X
Boris Cherny (Claude Code creator) covers practical mastery in ~28 min: Install via npm. Optimize setup… Ask codebase questions first… Use CLAUDE.md hierarchy…
— Grok (@grok) November 5, 2026 · View on X
What Grok says
Grok summarized Cherny’s practical masterclass arc — install, setup, ask the codebase like a senior engineer before editing, and invest in CLAUDE.md hierarchy across user and project scopes. That matches Anthropic’s documented memory model more than any single viral screenshot.
Critical analysis
The tutorial wave is healthy when it points at verification and memory; it is noise when it sells “10 minutes, no code” without staging, tests or review bandwidth. Anthropic’s own frontier workflow assumes AI-fluent teams, safe environments and willingness to red-team auto mode. Most companies should start with one triggered routine, a maker/checker split and a thin CLAUDE.md — then expand. Cherny’s delete-every-six-months line is the underrated point: memory compounds only if it is curated. Stale rules are technical debt with a friendly filename.
What’s still open
- Which community blueprints survive contact with messy monorepos and flaky CI?
- How far will Anthropic productize overnight subagent + verifier templates?
- Will enterprises standardize on agent-view-style fleet UIs across vendors?
- Where is the line between helpful auto mode and unaudited write access?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on Anthropic Claude Code documentation themes as summarized in public explainers (notably The Neuron’s Cherny/Wu breakdown), developer posts on X dated November 3–6, 2026, and Grok’s summary of Cherny’s practical talk. Tutorial checklists are community remixes, not an official Anthropic six-step product. More on Anthropic and Claude.
i10x Perspective
The coding-agent race is no longer about who can edit a file. It is about who can run a verified loop overnight and still understand what merged in the morning. Cherny’s playbook — and the tutorial economy around it — marks that shift. Copy the verification and memory discipline; leave the “army of a thousand agents” cosplay until your review queue can survive Tuesday.
