Claude Code’s Creator Does This Before Every Single Project

This video interviews Boris from Anthropic about practical practices for using Claude Code effectively. He explains verification-driven workflows, the role of claude.md files to capture project conventions and prohibitions, permission controls and modes (plan, auto-accept), and how Claude Code can orchestrate non-code workflows. Boris shares scaling habits — parallel sessions, background agents, cloud sessions and GitHub actions — and advocates using higher-quality models and test-driven verification to reduce errors and speed development. He also emphasizes lightweight, shareable commands and subagents to automate repetitive tasks.

– Verification-first workflows: always ask Claude to propose or run tests and show how it will verify output; testing creates a feedback loop that improves results.

– claude.md as project memory: keep a concise, repo-specific claude.md with project structure, style conventions and explicit “do not” rules; update it collaboratively so the agent avoids repeating mistakes.

– Permissioned automation & modes: use plan mode to design tasks, run the agent only after validating the plan, and set explicit permission rules instead of using “dangerously skip permissions” in production.

– Scale and orchestration: run multiple parallel sessions, leverage background/cloud agents and GitHub actions, and treat Claude like a junior developer—plan, verify, then execute.

Quotes:

Always give Claude a way to verify its work — that feedback loop instantly increases quality.

Claude Code is badly named: it’s not just for code — it orchestrates full workflows from Slack to BigQuery.

Treat the agent like a junior developer: plan in ‘plan mode’, verify, then run in auto-accept.