From Prompts to Loops: Why Claude Code’s Creator Deleted His IDE

Why the IDE Was Removed

Boris Cherny stopped opening his IDE after a month and uninstalled it, noting that during that time he submitted 259 pull requests, none of which contained a line of his own code—everything was generated by Claude Code. He announced that he would no longer write prompts, focusing solely on writing loops.

What Loop Engineering Is

Loop engineering means building a small system that automatically finds work, dispatches tasks, checks completion, and decides the next step. Addy Osmani coined the term “Loop Engineering” and described it as replacing the person who prompts the AI with a system that does the prompting.

Prompts Are Not Dead, They’re Inside Loops

The shift does not discard prompts; they are embedded in the new /loop and /goal commands. /loop repeats a block of prompt text at a set interval, while /goal runs until a user‑defined condition becomes true. After each iteration, a separate lightweight model judges whether the task is finished, preventing the generating model from self‑scoring.

Developers Have Been Using Loops Implicitly

Many users already employ loops without naming them. For example, anyone who has set a scheduled task in Claude Cowork has effectively written a loop. Simon Willison discussed “designing agent loops” as early as September 2025, showing that the concept predates the current hype.

The Three‑Layer “Hive” Architecture

Av1dlive reconstructed Boris’s configuration, calling it “THE HIVE.” It consists of three layers:

Local loop layer using /loop, which runs while a session is active (minimum once per minute).

Cloud‑based routines (Anthropic Routines) that execute hourly on cloned repositories, independent of the local machine.

Cluster layer using /batch and dynamic workflows to fan out work to hundreds of isolated worktree agents that run in parallel.

Seven ready‑to‑copy slash commands illustrate the first layer: monitoring PRs, mining Slack feedback, cleaning stale PRs, categorizing issues, persisting corrected rules into CLAUDE.md, etc.

Core Components of a Loop

Osmani’s blog breaks a complete loop into six components:

Automations : timed triggers that discover work and push it into an inbox.

Worktrees : isolated checkouts so multiple agents don’t clash on the same file.

Skills : a SKILL.md that stores project conventions and avoids the “goldfish memory” problem.

Connectors : MCP‑based bridges that let agents read issues, query databases, or post to Slack.

Sub‑agents : separate models that critique the primary coding agent’s output.

Memory : a simple markdown file that records what was done, what remains, and the next goal.

Not a Brand‑New Technology

These pieces echo earlier work such as ReAct (2022), AutoGPT (2023), and Geoffrey Huntley’s “Ralph loop” (July 2025). Both Claude Code and Codex have implemented almost identical component sets, so the real competition now hinges on who makes loop authoring the smoothest.

Key Challenges

The approach faces three major hurdles:

Cost : stacking multiple agents and sub‑agents on top‑tier models can consume 5‑10× the expected token budget. Steinberger jokes he has “unlimited tokens,” a privilege not shared by most users.

Loss of Control : Anthropic warns of three pitfalls—lazy shortcuts (e.g., checking only 20 of 50 security items), self‑praise (agents rating their own work highly), and drift (a “don’t do X” rule disappearing after many iterations).

Skepticism : Critics argue that loop engineering merely rebrands existing agent‑loop or task‑scheduling ideas, offering no fundamentally new capability.

Practical Advice

Osmani cautions that stronger loops increase three risks:

Verification still requires human oversight; a loop’s “completed” flag is not proof of correctness.

Comprehension debt grows as loop‑generated code accumulates, making it harder to understand later changes.

Using loops as a shortcut to avoid understanding the underlying problem can accelerate technical debt.

Conclusion

Boris’s confidence to delete his IDE stems from moving prompts into systematic loops rather than abandoning them. Developers are encouraged to start building their own loops, but they must remain the engineers who understand what the loops are doing, not just the button‑pressers.