Originally published at claudeguide.io/how-anthropic-employees-use-claude
How Anthropic Employees Use Claude (Patterns Worth Stealing)
Across publicly shared Anthropic posts, podcasts, and engineer interviews from 2024-2026, ten patterns appear repeatedly: writing CLAUDE.md as the source of truth, using plan mode for anything multi-file, deferring to subagents for parallel research, encoding repetitive workflows as skills, treating Claude as a code reviewer (not just a writer), aggressive prompt caching, structured output for data tasks, scheduled tasks for monitoring, the "ask Claude to code-review your prompt" loop, and the discipline of small, well-scoped tasks (10 patterns total). None of these are secret techniques — they're public-facing best practices, but they appear so consistently across Anthropic engineering content that they're clearly the dominant internal style.
Sources for this article
This article is compiled exclusively from public sources. No internal Anthropic information is included. Public sources include:
- The Anthropic engineering blog
- Public talks at conferences (LangChain, AI Engineer Summit, etc.)
- Podcast appearances (Lenny's Newsletter, Latent Space, No Priors)
- Engineer interviews (TechCrunch, The Information, Wired)
- Anthropic Cookbook and example repositories on GitHub
- Posts on the Anthropic Discord that engineers have explicitly made public
Where a specific person is referenced, they're someone who has spoken publicly under their own name about their workflows.
Pattern 1: CLAUDE.md as the project's source of truth
The single most consistent pattern across Anthropic-shared workflows is that every active codebase has a well-maintained CLAUDE.md file. It's not an afterthought — it's the first file new contributors read.
What's in a typical Anthropic CLAUDE.md (based on public templates and shared examples):
- Project context: what does this codebase do, in 3–5 sentences
- Development setup: install commands, environment variables, common gotchas
- Code style: framework conventions, naming patterns, testing approach
- Workflows: how to run tests, deploy, debug common issues
- Don't-do list: anti-patterns specific to this project
- File map: where the important things live
Why this matters: When Claude reads CLAUDE.md before any task, it inherits the project's context for free. Without CLAUDE.md, Claude infers context from file structure — which is fine for small tasks but degrades on multi-file work.
Pattern to steal: maintain CLAUDE.md as actively as you maintain README. Update it when conventions change. Treat outdated CLAUDE.md as a bug.
Pattern 2: Plan mode for anything that touches
→ Get Claude Code Mastery — $29
30-day money-back guarantee. Instant download.
Sources
- Anthropic Engineering Blog
- Building Effective Agents — Anthropic
- Mike Krieger on Lenny's Podcast (2025)
- Anthropic Cookbook (GitHub)
- Claude Code documentation
- AI Engineer Summit 2025 talks (Anthropic engineers)
Top comments (0)