Reflecting on a Year of Claude Code — Boris Cherny & Cat Wu (2025)

A retrospective conversation between Boris Cherny (Head of Claude Code) and Cat Wu (Head of Product, Claude Code) marking one year since Claude Code's general availability. They trace the product's evolution from a two-reaction Slack demo to an agentic platform running thousands of parallel agents, covering verification, auto mode, routines, context minimalism, and the convergence of engineering and product roles.

Key Takeaways

Persisting lessons from Claude mistakes into CLAUDE.md or skills — rather than ad-hoc corrections — enables Claude to "run forever" 0:55
Verification for agents is not unit tests or linting; it is enabling the agent to run and self-test the thing it built 1:24
Auto mode is safer than permission-prompt-per-action because humans rubber-stamp 99% of prompts anyway, causing attention to glaze over 8:07
The progression is: source code → talking to an agent → talking to a loop or routine that prompts the agent 10:53
"Context minimalism" has superseded prompt engineering and context engineering: give the model the minimal system prompt and minimal tools, then let it pull in context itself 16:29
The PM/engineer boundary is dissolving; designers, finance, and data teams at Anthropic now run Claude Code routinely 4:06

Verification: Beyond Unit Tests

Boris distinguishes agent verification from conventional automation 1:24: linting and type-checking were already automated; what matters for agents is whether the agent can actually run the thing it built. Working examples from the team:

With Claude Opus 4, Claude was prompted to "build a feature, then test yourself in bash," whereupon it opened a Claude CLI and self-tested — at the time surprising, now standard 1:50
Loops now cover iOS simulator, Android simulator, and desktop via computer use 2:15
Cat's desktop development skill teaches Claude to spin up the local app, use computer use to click through new UX, test edge cases, fix issues, and recheck 2:56
When staging is flaky, Claude reads Slack to check whether others have already hit the issue before debugging 2:43

Auto Mode: Design, Safety, and Red-Teaming

Boris describes plan mode as obsolete for newer models: starting with Claude Opus 4.6 and definitively with 4.7, the planning step is no longer needed 7:01. He now uses auto mode exclusively.

Cat explains the security rationale 8:07: when users accept 99% of permission prompts, attention degrades on the 1% that matter — auto mode routes suspicious requests to a classifier instead of the user. Steps taken before shipping auto mode 9:03:

Collected thousands of agent-trajectory transcripts, had the classifier label safety — achieved high accuracy.
Hired external red teamers to attempt prompt injection and codebase exploits.
Built evals from the red-team outputs.
Had internal teams attempt additional attacks; iterated auto mode until all were blocked.

Boris's stance: many safety features initially seemed implausible ("route the prompt to a model? No way") but empirically proved correct — a recurring lesson about building on top of ML models 9:51.

Routines and Loops: The Next Abstraction Layer

Cat describes routines as the first obvious application of the Claude Code SDK 6:21:

One engineer set up a routine that listens for every GitHub issue and bug report about voice mode, proactively raises a fix PR, and pings the engineer 4:53
A second routine watches for bug reports unresponded to within 5 hours and auto-proposes fixes 5:47
Result: Cat's own bugs were being fixed by another engineer's Claude before Cat's Claude even flagged them 5:36

Boris frames the conceptual leap 10:32: the industry went from writing source code → talking to an agent that writes source code → now talking to a loop/routine that prompts the agent. Two paradigm shifts in roughly 18 months.

Role Convergence and Organizational Change

Both speakers observe that engineers adopt Claude Code first; adjacent roles — designers, PMs, finance, data science — follow by watching over engineers' shoulders 4:06. Within the Claude Code team specifically 13:43:

Product team all writes code
DevRel team all writes code
Design team all writes code
Engineers ship products end-to-end including legal, marketing, and security coordination

Boris draws an analogy to a Harvard Business Review case study from the 1990s on why productivity gains from personal computers lagged 11:13: companies tried to keep paper processes and add computers on the side. The payoff came only when the computer went to the center of every business process — the same transformation AI requires. He notes the AI transition is faster because work is already digitized and Claude can use a computer and run code 12:43.

On the product-vs-engineering question: Boris and Cat jointly predict everyone will be both 13:36, with AI disproportionately benefiting people who combine curiosity, product taste, and appetite for end-to-end ownership.

Multi-Agent Workflows and Context Minimalism

Boris's current workflow for managing hundreds of agents 14:40:

Single terminal tab using the new agent view instead of six checkout tabs
Desktop app handles git worktree cloning automatically
Roughly half of engineering done from phone via Remote Control; starts agents on laptop, checks in and spawns new agents while mobile
Uses voice mode to start new agents on the spot mid-conversation

On context strategy 16:13: Boris argues that prompt engineering matched the Sonnet 3.5 era; context engineering matched the Opus 4 era; today's models need neither. The prescription is a minimal system prompt, minimal tools, with a mechanism for the model to pull in context on demand. Cat echoes this as "context minimalism" — over-specifying context is micromanagement that prevents the model from finding a better solution 16:43. The team is also slimming its own harness to leave more room for user prompts 17:09.