Boris Cherny, the Anthropic engineer who created Claude Code, has publicly declared the end of manual prompting for his own workflow — and his colleagues are following him. In a string of interviews and a viral video circulating on X this week with more than 870 posts in eleven hours, Cherny says he now spends his day writing loops that prompt Claude on his behalf, while hundreds of agents work through the night on his behalf.
Key Highlights
- Boris Cherny: "I don't prompt Claude anymore. I write loops, and the loops do the work. My job is to write loops."
- 100 percent of pull requests at Anthropic are now run through Claude Code, according to Cherny.
- Anthropic recently deleted roughly half of Claude Code's system prompt after the latest model upgrades, signalling that newer models need less hand-holding.
- Cherny personally runs five terminal instances of Claude Code in parallel, plus five to ten additional sessions on claude.ai/code, with recurring loops such as
babysit-prs,post-merge-sweeper, andpr-prunertriggered on five, thirty, and sixty minute intervals. - He says nightly orchestration regularly runs "hundreds, sometimes thousands of agents" for between five and twenty hours each.
Details
The shift is being formalised under a new term, Loop Engineering — a workflow in which the developer's primary output is not code or even prompts, but the agentic loop that decides when, how, and on what an AI coding agent should work next. In a recent interview with The New Stack, Cherny described the discipline as the natural evolution of prompt engineering once models become reliable enough to self-verify.
Cherny's January 2026 thread describing his workflow went viral, racking up millions of views and inspiring howborisusesclaudecode.com, a community-maintained reference of his patterns. Among the documented practices: he uses Opus 4.5 and later (4.7 and 4.8) with thinking mode on every task, defaults to xhigh reasoning effort, and reserves max effort for the hardest debugging sessions. Rather than correcting Claude in chat, he writes mistakes into a shared CLAUDE.md file checked into git, so every future run benefits from the correction — a compounding error-reduction pattern his team calls memory through repos.
The /loop slash command, paired with /goal for stop conditions, is now the centrepiece. Typical recurring loops at Anthropic include:
5m /babysit— auto-fix failing builds and address review comments on open PRs.30m /slack-feedback— open new PRs based on feedback dropped in team Slack channels.1h /pr-pruner— close stale or duplicate pull requests automatically.
Impact
The productivity claims are striking. Cherny told Platformer that Claude Code itself "has been 100 percent written by Claude Code for over six months," and that internal Anthropic data shows engineers shipping roughly 70 percent more per head since the loop-driven workflow became standard. Anthropic CEO Dario Amodei has separately stated that around 90 percent of Claude's own product code is now AI-written.
For the rest of the industry, the implication is that the bottleneck is shifting upstream. Writing the next prompt is no longer the scarce skill — designing a well-bounded loop with clear verification gates, idempotent steps, and a safe stop condition is. Cherny argues that this changes the role itself: "I think we're going to start to see the title 'software engineer' go away. And I think it's just going to be maybe builder, maybe product manager."
Background
Anthropic released Claude Code in early 2025 as a terminal-first agentic coding tool. It quickly became the fastest-growing AI coding product on the market, sustained by features such as plan mode, MCP integrations, headless cron via /loop, custom subagents, and a managed cloud variant that runs sessions in isolated sandboxes. Earlier this year, Anthropic added orchestrated parallel subagents under a research-preview workflow API, the foundation for the multi-agent loops Cherny now relies on.
The shift mirrors a broader pattern across the frontier labs. OpenAI's Codex CLI and GitHub Copilot have introduced their own asynchronous, long-running agent modes in recent months, and Cursor shipped Cloud Agents earlier this year. Verification loops, git worktrees, and overnight runs are increasingly the default rather than the exception for engineers working on large codebases.
What's Next
Cherny has hinted that Anthropic will continue to invest in loop primitives, with deeper integration between /loop, /goal, custom skills, and the new dynamic workflow API. Expect more first-class support for verification harnesses — bash, tests, simulators, browser drivers — built directly into Claude Code, since Cherny credits verification with a two to three times quality boost on autonomous runs.
For MENA software teams, where engineering capacity is often the bottleneck on ambitious roadmaps, Loop Engineering offers a path to compress the typical PR cycle from days to hours. The tooling is already mature enough to adopt today — what remains is the cultural shift from babysitting agents to designing the loops that babysit themselves.
Source: The New Stack — Loop Engineering