Startup School

Boris Cherny on treating the model like a living creature

Boris Cherny· Creator of Claude Code at Anthropic
·~36 min·English·Y Combinator
AgentsLLMAI SafetyReasoning
TL;DR

Boris Cherny, creator of Claude Code, argues that building on frontier models is an empirical craft: delete your scaffolding every generation, unhobble capability the product is hiding, and let the model verify its own work while it runs for days.

01Core Mental Model

Treat the Model Like a Living Creature

Building on a model is not systems engineering, it is animal husbandry: every generation has a different personality, so you get to know it and adjust the harness, rather than designing everything up front.

every model generation it behaves differently. it has a slightly different personality and you have to take the time to get to know it and then adjust the harness based on that

— Boris Cherny, Startup School
Key Insight
The old craft rewarded upfront design and stable abstractions; this one punishes them. When the substrate mutates every few months, the only durable skill is the scientific loop itself, not any particular architecture.
02The Core Practice

Press Delete Every Six Months

Each new model release begins by deleting the system prompt and rebuilding it line by line — for Opus 5 that meant cutting 80%, and the same delete-first discipline applies to your CLAUDE.md, skills, and hooks.

you delete the entire system prompt and then you bring it back line by line to figure out what is the impact of each individual line

— Boris Cherny, Startup School
Key Insight
Most prompt bloat is scar tissue: instructions that corrected a weakness in an older model and were never removed. A smarter model turns yesterday's guardrails into dead weight it still has to read on every single call.
03The Builder's Opening

Unhobble the Model, Capture the Overhang

Today's models can already do far more than any product lets them, and that gap is the opportunity — Claude Code was born by stripping the scaffolding off Sonnet 3.5 so it could write whole files while rivals shipped single-line autocomplete.

the model can do this at every given model generation, but there is often not a product that lets the model do this and lets it express this kind of ability to do this

— Boris Cherny, Startup School
Key Insight
Cherny is effectively telling a room of founders that the next Claude Code already exists inside today's models. The scarce resource is not model capability, it is the imagination to build the product that lets that capability out.
04Proof of Capability

The Eleven-Day Rewrite

A single dynamic-workflow prompt rewrote the entire Bun runtime from Zig to Rust in 11 days — over 100,000 lines, verified against Bun's own test suite, and now running in production inside Claude Code.

And it ran for 11 days and it rewrote the entire codebase

— Boris Cherny, Startup School
Key Insight
The rewrite worked because Bun was exhaustively tested: the test suite was the oracle that let the model check itself for eleven days without a human in the loop. Capability plus a verifier is what turns a demo into production code.
05The Skill People Miss

Verification Is the Whole Game

The single skill people get wrong is giving the model a way to check its own work — hand it a task that is slightly too hard, wire up a verifier, and it will grind for days without getting stuck.

the verification I think is probably the single most important thing that people do not get right

— Boris Cherny, Startup School
Key Insight
Cherny's own Electron-to-Swift experiment worked because the prompt built in the check: run the app, screenshot it, compare pixel by pixel, do not stop until done. Verification is not QA at the end, it is the leash that keeps a long run on task.
06The Frontier Shift

The Lethal Trifecta, Defused

Three stacked defenses have made prompt injection something Anthropic can no longer demonstrate: a well-aligned model, a prompt-injection classifier that runs across all traffic and watches the neurons that light up, and a second auto-mode classifier on top.

with these three layers, we just cannot demonstrate prompt injection anymore

— Boris Cherny, Startup School
Key Insight
Defense in depth is the tell here: no single layer is trusted. The most striking is the interpretability classifier, which reads the model's internal state rather than its output, catching an attack the model itself would not report.
07How the Scale Works

An Algebra for Agents

Dynamic workflows let one prompt orchestrate thousands of agents in sequence and in parallel — built by a functional-programming mind, it behaves like an algebra and opens a new place to spend test-time compute.

it's essentially an algebra for agents. So there's a way to run agents in sequence. There's a way to run agents in parallel

— Boris Cherny, Startup School
Key Insight
Framing orchestration as an algebra reframes scaling itself. Test-time compute stops being just how many tokens one model generates and becomes how you compose many runs, which is a lever product builders can pull without touching the weights.
08Where This Is Going

The Codebase That Maintains Itself

Twenty to thirty routines now run every day across Anthropic's codebases — deleting dead code, unifying duplicate abstractions, and managing test coverage — doing the maintenance that used to take dozens or hundreds of engineers.

This is again hundreds of agents running every day, sometimes thousands of agents every day. It's doing the work of, you know, dozens or hundreds of engineers

— Boris Cherny, Startup School
Key Insight
The point is not that machines replace maintainers, it is what the freed humans do instead: ship products and talk to users. When upkeep automates itself, the scarce human contribution shifts decisively from maintenance to judgment and taste.