Boris Cherny

Overview

Boris Cherny is an AI coding and developer-experience commentator and product/operator voice most closely associated here with Claude Code, code-review workflows, and practical discussion of failure modes in LLM-powered software. Across newsletter mentions, he appears as a recurring source on how coding agents should be evaluated, how developer tooling should expose usage and limits, and how teams can operationalize safer and more effective AI-assisted software development.

For AI Product Managers, Boris Cherny matters because his commentary sits at the intersection of product design, model behavior, tooling ergonomics, and real-world deployment risk. His notes span lightweight but high-leverage product features such as `/usage` and `/checkup`, model-evaluation claims such as `/code-review low` outperforming alternatives at very low cost, and broader framing on where LLM bugs are moving next: away from narrow syntax mistakes and toward system design flaws, UX issues, and missing context. That combination makes his perspective especially useful for PMs building agentic coding, review, and developer-platform products.

Key Developments

• 2026-05-11: Boris Cherny launched a new `/usage` command in Claude Code that gives a detailed breakdown of what actions are consuming tokens and what limits apply, helping users self-serve debugging. He also noted that 5-hour run limits were doubled across Claude Code plans.
• 2026-05-14: He said Mythos Preview was the first model to fully solve UK AISI cyber ranges end-to-end, including the previously unsolved “Cooling Tower,” and said it was being deployed to defenders quickly but responsibly, with more to come on Glasswing.
• 2026-05-22: He rolled out a `/usage` command in Claude Code’s CLI to break down token usage by Skills, Agents, MCPs, and Plugins, with desktop support planned next.
• 2026-06-29: He described five archetypes on the Claude Code team—Prototyper, Builder, Sweeper, Grower, and Maintainer—highlighting how roles increasingly blend product, engineering, and design.
• 2026-07-09: He rolled out `/checkup` in Claude Code to automate cleanup of unused skills, MCPs, and plugins, and to dedupe and split CLAUDE configuration.
• 2026-07-24: He described using Fable’s dynamic workflows plus profiling to iteratively tune code until p95 latency falls below 300 ms, illustrating a performance-first workflow for AI-assisted development.
• 2026-08-08: He said that combining model training, input probes, and an intent-checking classifier can reduce indirect prompt injection on unseen attacks to near zero, and announced that auto mode would become the default in Claude Code the following week.
• 2026-08-09: He commented that evaluation harnesses can support other models through proxies such as LiteLLM, but that effective harnesses still require substantial model-specific tool design, prompting, and tuning. He also raised a UX question about whether Claude should automatically resume using Fable after a user’s limit resets.
• 2026-08-10: He shared a prompt-injection benchmark from an independent researcher and said Anthropic had largely solved the threat in practice for Claude models, with red-team testing showing similar results. He framed prompt injection as a common attack path against people and agents, including malicious webpage text designed to exfiltrate credentials.
• 2026-08-12: He said unspecified evaluations found `/code-review low` produced better results than other models at a fraction of the cost, reportedly under $0.01. He also argued that LLM bugs are shifting from off-by-one errors toward system design, UI usability, and lack of broader context, recommending adversarial code review—such as edge-case prompts in an iOS simulator or Claude’s built-in `/code-review`—to catch many issues.

Relevance to AI PMs

1. Design evaluation around real workflows, not just model benchmarks. Boris Cherny’s comments on `/code-review`, harness quality, and adversarial testing suggest PMs should evaluate coding agents in realistic environments with tools, simulators, and edge cases—not just offline benchmark scores.

2. Make costs and limits legible inside the product. His emphasis on `/usage` shows the value of exposing token consumption by feature layer—agents, skills, plugins, MCPs—so customers can debug spend, understand behavior, and trust the system.

3. Prioritize safety and UX together. His framing of prompt injection defenses and shifting LLM bug categories is a practical reminder that the next generation of failures often comes from workflow design, context gaps, and unintended agent actions, not only raw model mistakes. PMs should treat safety, review, and recovery controls as core product features.

Related

• Anthropic / Claude / Claude Code: The primary ecosystem in which Boris Cherny is mentioned, especially around developer tooling, code review, usage visibility, and safety.
• claude-code-review / code-review: Central to his commentary on low-cost evaluation quality and adversarial review workflows.
• prompt-injection: A major theme in his safety comments, including layered mitigations, benchmarks, and red-team validation.
• LiteLLM: Referenced as a proxy layer for supporting non-native models in evaluation harnesses, though Cherny notes strong model-specific adaptation is still required.
• Fable: Connected through workflow automation, latency tuning, and UX questions around automatic resumption after limits reset.
• Mythos Preview / UK AISI / Glasswing: Tied to his comments on cyber-range performance and deployment to defenders.
• Prototyper / Builder / Sweeper / Grower / Maintainer: Team archetypes he used to describe how AI product teams are organizing across blended product, design, and engineering responsibilities.
• iOS simulator: Part of his recommended adversarial testing toolkit for catching edge-case product bugs in AI-generated software.
• Bedrock / Vertex / Azure / GPT-4 Turbo / Opus / Sonnet 4.6: Relevant as comparison or deployment context for cross-model evaluation, even where his comments stress that harness quality and tool integration matter as much as raw model access.