CLAUDE.md as Cultural Object — The Karpathy Moment #
- forrestchang/andrej-karpathy-skills (GitHub) — A single CLAUDE.md file distilled from Andrej Karpathy’s January 26, 2026 X posts about LLM coding pitfalls. Four principles: Think Before Coding (state assumptions, ask rather than guess); Simplicity First (minimum code, no speculative abstractions); Surgical Changes (touch only what’s necessary, match existing style); Goal-Driven Execution (give success criteria, not instructions). Karpathy’s framing: “Don’t tell it what to do, give it success criteria and watch it go.” The repo reached 137K stars — one of GitHub’s fastest trajectories. Distilled by Forrest Chang; principles from Karpathy’s X thread.
- Karpathy CLAUDE.md: The 65-Line File With 100K GitHub Stars (Miraflow) — Detailed account: hit #2 on GitHub trending with 5,828 stars in a single day (April 13, 2026). Reported accuracy improvement from 65–70% to 91–94%. Karpathy’s own framing: his coding shifted from “80% manual+autocomplete, 20% agents” in November 2025 to “80% agent coding, 20% edits” by December 2025. Community response: recognition of articulated frustrations, not discovery of novel concepts.
- Karpathy-Inspired CLAUDE.md: How to Add It to Any Project in 30 Seconds (Alpha Signal, Substack) — Practical installation guide; also notes adoption via the Claude Code plugin marketplace as a drop-in.
- Karpathy CLAUDE.md Skills: Use the Viral Rules as a Menu, Not a Template (Developers Digest) — Practitioner pushback on blind adoption: the four principles are starting points, not copy-paste rules. Key tension: Surgical Changes and Simplicity First can conflict when a surgical fix produces messy code — judgment required.
- shanraisshan/claude-code-best-practice (GitHub) — “From vibe coding to agentic engineering.” Community-maintained CLAUDE.md and workflow templates for professional-grade agentic setups; companion to the Karpathy file rather than derivative.
- abhishekray07/claude-md-templates (GitHub) — CLAUDE.md templates collection covering role-specific and stack-specific variants — indicative of the template ecosystem Karpathy’s moment catalysed.
Hooks — 27 Events, 5 Handler Types #
- Claude Code Hooks: The Complete 2026 Production Reference (The Prompt Shelf) — As of v2.1.141+: 27 distinct events (32+ subtypes via matchers like
startup, resume, clear, compact on SessionStart). Five handler types: command (shell script), http (webhook POST), mcp_tool (delegates to MCP server), prompt (single-turn LLM gate), agent (full subagent with multi-turn reasoning and tool access). The agent handler suits complex investigations (diagnosing failures, deep analysis) but is the heaviest option. Critical: only exit code 2 blocks an action — exit code 1 is non-blocking. - Claude Code Hooks: Automate Your Coding Workflow in 2026 (Kjetil Furas) — Practitioner walkthrough covering the expanded event taxonomy and agent handler in production.
Agent SDK — June 15 Billing Split & Model Deprecations #
- Claude Agent SDK Changes June 15, 2026: Migration Playbook (ThePlanetTools) — Operational alert: from June 15, Agent SDK programmatic usage (Python/TypeScript SDKs,
claude -p, GitHub Actions, third-party apps) moves to a separate monthly credit ($20 Pro / $100 Max 5x / $200 Max 20x). Interactive Claude Code sessions remain on plan quota. Two model IDs retire: claude-sonnet-4-20250514 → migrate to claude-sonnet-4-6-20260217; claude-opus-4-20250514 → migrate to claude-opus-4-7. API calls using deprecated model IDs return errors after June 15. - Anthropic’s June 15 Billing Change: What Every Claude Code & Agent SDK User Must Do (Coders Era) — Migration checklist: audit hardcoded model IDs, update to Sonnet 4.6 / Opus 4.7, tag SDK workloads separately in cost dashboards, enable billing alerts at 50% and 80%.
OpenClaw — Self-Hosted Agent Runtime #
- OpenClaw vs Claude Code Channels vs Managed Agents: Which Should You Use in 2026? (MindStudio) — OpenClaw is a new open-source agent runtime you self-host, with data staying entirely in your own environment — suited for regulated industries (finance, healthcare, government) where data residency is non-negotiable. Three-way positioning: OpenClaw (self-hosted, data control) vs Claude Code Channels (purpose-built dev workflows, IDE/CI) vs Managed Agents (fully hosted, zero infra). First appearance of OpenClaw in this journal.
Boris Cherny — How the Head of Claude Code Actually Works #
- How Boris Uses Claude Code (howborisusesclaudecode.com) — Boris Cherny’s (Head of Claude Code, Anthropic) documented personal workflow: 5 parallel Claude instances across 5 separate git checkouts. Plan mode → one-shot implementation. Ships 20–30 PRs/day. The practitioner-as-benchmark form — showing the ceiling of what’s achievable rather than teaching beginners — is more useful as a calibration tool than as a tutorial.
- Building Claude Code with Boris Cherny (Pragmatic Engineer) — Orosz interviews Cherny on origins and architecture: Claude Code evolved from a terminal prototype to 4% of public GitHub commits. Cherny’s view: the 4% figure underestimates impact because it misses PRs that AI shaped without authoring.
- Head of Claude Code: What Happens After Coding Is Solved (Lenny’s Newsletter) — Cherny on the trajectory beyond AI-solved coding: the engineering role’s evolution, and broader implications for software development as a profession.
CLAUDE.md Architecture — The Compliance Budget #
- Designing CLAUDE.md Correctly: The 2026 Architecture (ObviousWorks) — References Boris Cherny’s ~2,500-token (~100 line) internal CLAUDE.md at Anthropic and the ~150–200 instruction compliance budget before adherence drops. The key design principle: CLAUDE.md is advisory (~80% compliance); hooks are deterministic (100%). Design for hooks where 100% matters; reserve CLAUDE.md for guidance where some drift is acceptable.
Routines — The Third Execution Mode #
Hooks — Production Implementation Guides #
Cross-links #
- Vibe Coding Approaches Karpathy’s coding-mode shift (80% agent by December 2025) is a concrete, precisely documented data point for the “agentic engineering” transition.
- Open vs Closed AI Ecosystems OpenClaw as a self-hosted alternative to Managed Agents is the open-source response to VentureBeat’s lock-in warning from the May 2 gather — data residency requirements are the concrete forcing function.
- Claude Integrations The Agent SDK billing split (June 15) separates “developer tool” from “platform infrastructure” at the billing layer — a quiet but significant signal about how Anthropic is segmenting the two use cases.
- Vibe Coding Approaches Boris Cherny’s 20–30 PRs/day workflow is the current ceiling benchmark for what agentic engineering looks like in practice.
- Emerging pattern: CLAUDE.md has become a cultural artefact, not just a config file. The Karpathy repo is one of GitHub’s fastest-growing ever. The community now treats CLAUDE.md authoring as a first-class skill with visible exemplars, templates, and derivative discourse.
- Emerging pattern: The “practitioner-as-CLAUDE.md-brand” form — Forrest Chang distilling Karpathy, Boris Cherny’s tips-as-skill — is repeating. Watch for named practitioners who build followings around CLAUDE.md configurations.
- Emerging theme: The June 15 billing split (SDK vs interactive) is the first time Anthropic has drawn a pricing line between developer tool and programmatic agent infrastructure. If this persists, it signals Managed Agents and Claude Code are converging toward different market segments, not just different use cases.
- Quality signal: The CLAUDE.md compliance budget (~150–200 instructions before adherence drops) is the most concrete design constraint in this gather cycle — it turns CLAUDE.md authoring from an art into an engineering problem.
- Keyword suggestion:
"CLAUDE.md" example OR template OR "inspired by" -beginner — captures the active CLAUDE.md exemplar discourse distinct from generic “write a CLAUDE.md” guides. - Author to watch: Forrest Chang (forrestchang, GitHub) — distilled Karpathy and triggered the viral moment; likely to produce more high-signal CLAUDE.md work.
- Source to watch: thepromptshelf.dev — produced the most complete hooks reference found; matches preferred-source quality.
- Gap: No coverage of how teams are handling CLAUDE.md versioning (git-tracked vs per-developer) as “team CLAUDE.md” patterns emerge alongside individual ones.