Loops, Not Prompts

Boris Cherny on Claude Code at Anthropic — the next abstraction layer in programming

← Boris Cherny hub · 4 Steps of AI Adoption · AI Agents

Key insight: The highest-leverage users of AI coding tools are no longer writing prompts directly. They are writing loops — autonomous systems that prompt models, orchestrate exploration, verify output, and iterate. This is the signal of where power users and agentic systems are heading.

This page summarizes the Acquired Podcast conversation with Boris Cherny (creator of Claude Code, formerly "Cloud Code"). Reference the video / X post here.

Origin Story

Boris joined Anthropic's prototyping / Labs team in late 2024. The mandate was clear: build the next big product while continuing to push model frontiers.

At the time, early coding tools were mostly limited to autocomplete or basic Q&A. The team went all-in on a full coding agent — what became Claude Code.

It started rough. Initially it only handled ~10-20% of Boris's code. Rapid improvement came from better base models (Sonnet 4 / Opus 4 → Opus 4.5), harness features, multiple UI surfaces (CLI → desktop app → mobile → Slack / GitHub apps), and powerful new modes like Plan mode.

Anthropic's core mission is AI safety. Coding turned out to be an ideal "Petri dish" for studying models in the wild: clear pass/fail criteria (does it compile? run? pass tests?), constrained solution spaces, and rich real-world interaction data. This kind of grounded, high-stakes environment is valuable for safety research beyond traditional interpretability work in the lab.

The Evolution: From Prompting to Writing Loops

Boris uninstalled his IDE months ago. He no longer writes much code manually.

The bigger shift: He stopped directly prompting Claude for most work. Instead, he writes loops — autonomous systems that:

Prompt Claude (or other models)
Orchestrate multi-step tasks
Explore the codebase or problem space
Iterate, verify, and self-correct

His job evolved into "writing loops." This is the next abstraction layer in the long arc of programming:

punch cards → assembly → high-level languages → AI agents & agentic loops.

The builder's edge is no longer raw prompting skill. It is orchestration, verification, feedback design, and encoding principles ("skills" for models).

Impact at Anthropic

The productivity numbers are striking (and the early 3X+ figure is now considered outdated — the real multiplier is likely higher):

Code output per engineer has increased dramatically even as the team grew.
New hires ramp up in ~2 days instead of weeks. They simply query Claude about the existing codebase.
Everyone uses the tools daily — not just engineers and researchers. Designers, finance, the chief of staff, and other non-engineers are shipping code.

Roles are collapsing into generalist "builders." One person handles scoping conversations with users, light design, data analysis, implementation, and shipping. The old functional silos (PM / designer / engineer) are dissolving in practice.

Claude Cowork and Organic Adoption

Claude Cowork is the GUI version aimed at non-engineers (no terminal required). It was built in roughly one week — using Claude Code itself.

Early experiments (Slack bots, browser extensions) often failed due to friction (e.g. no easy filesystem access). The team learned to prioritize building what they themselves love and use every day. Organic adoption by data scientists and even whimsical users (one person used it to automate tomato plant monitoring) validated the direction.

Advice for Companies and Founders

Maximum Tokens, Maximum Experimentation

Give everyone generous token budgets and let them play. Constraints kill the learning flywheel.

Deliberately Underfund Projects

Assign fewer people + heavy AI usage. The constraint forces automation and creates compounding efficiency gains.

Shift Budget from Headcount to Compute

Upfront costs rise, but the ongoing marginal cost of output drops dramatically.

Embrace Generalists

This is the golden age for versatile builders who can scope, design, analyze, implement, and ship.

Product Taste May Erode — Values Remain

As models get better at ideation and feedback analysis, the durable human role is teaching values: how to be "good" the way we teach children.

Key Themes

Coding is commercially valuable. It funds safety work without ads and is exactly what enterprise customers will pay for.
Exponential acceleration. Log-linear thinking is already outdated. Roles, tools, and organizations must adapt at the same pace as the models.
From chat to systems. The frontier has moved from manual prompting to agentic loops and autonomous processes. The new craft is designing the orchestrator, the verifier, and the skill library.