Boris Cherny: Building Claude Code

Why does it matter? Because the biggest AI opportunity right now is subtraction, not addition

Boris Cherny built Claude Code not by engineering elaborate prompts — but by deleting them. His framework of "product overhang" exposes a counterintuitive truth: most AI products actively cripple the models underneath them, and the biggest startup opportunity today requires no breakthrough — just getting out of the way.

• Today's models can do far more than any product allows — the constraint is the harness, not the model. • Every model release is an invitation to delete your system prompt and rebuild from observed failures only. • Verification — giving the model a way to check its work — matters more than any prompt engineering technique. • Opus 5's resistance to prompt injection makes autonomous agents in adversarial environments viable for the first time.

The biggest AI startup opportunity is already inside today's models — nobody's capturing it

"There is so much product overhang that I'm not seeing startups capture." Claude Code was born from exactly this: Sonnet 3.5 could write entire files at once, but competing products were still doing single-line autocomplete. The model was ready; the product was in the way. Cherny calls this suppression "hobbling" — and believes there are "dozens, hundreds" of undiscovered capabilities sitting inside current models. Someone gave Opus 5 OpenCV and asked it to draw portraits and landscapes. Nobody trained it for that. It just could.

Claude Code deleted 80% of its system prompt for Opus 5 — and the model got smarter

They deleted 80% of Claude Code's system prompt for Opus 5. The removed instructions weren't wrong — they were patches for behaviors the model has now internalized. Run it in simple mode and it actually performs better without them. Cherny's method is strictly empirical: delete everything, observe where the model repeatedly fails, add back only those specific lines. "You don't want to guess what the model needs because you might not predict it correctly." For your own CLAUDE.md, skills, and hooks: wipe them on the next release, then rebuild from what you actually observe failing.

Verification beats prompt engineering — it's what makes weeks-long autonomous tasks possible

The task still running after two weeks was launched with one paragraph: rewrite this Electron app in Swift, screenshot it, compare pixel-by-pixel to the original, don't stop. No scaffolding. Just a clear signal for success. The Bun-to-Rust rewrite ran on the same logic — a comprehensive test suite gave the model an unambiguous check on its work, so it ran for 11 days and rewrote a production JavaScript runtime. That would have taken "even with the best engineers, definitely over a year." Design the verification mechanism first. The rest follows.

Opus 5 doesn't appear to be prompt injectable — the classifier reads neurons, not outputs

"Opus 5 does not seem to be prompt injectable anymore." Three layers converged: three years of alignment training, an auto-mode classifier, and a mechanistic interpretability-based classifier that watches neuron activations directly — "we're literally looking at neurons in the model's brain that light up when prompt injection happens." Security researchers, red teams, and open competitions have all failed to demonstrate an attack. Agents browsing untrusted web content — previously the most dangerous deployment pattern — just became viable.

One prompt can now spawn thousands of agents running for weeks — a new form of test-time compute

One prompt, one dynamic workflow, 11 days: the entire Bun runtime rewritten from Zig to Rust, now shipping in production inside Claude Code. Dynamic workflows are "an algebra for agents" — sequential, parallel, fan-out, fan-in — staging thousands of coordinated agents with verification between passes. This expands what test-time compute can mean: not just tokens from a single model, but agents across a multi-stage, verified pipeline. The old definition of a prompt — one instruction to one model — is already obsolete.

Anthropic now has Claude running 20–30 daily routines to maintain its own codebase

Every routine is one sentence. "Clean up dead code." Claude independently chose to use static and dynamic analysis; nobody specified that. Other daily jobs: ship experiments that hit 100% rollout, write missing test coverage, delete useless tests left by older models, and "abstraction police" — finding the same abstraction reimplemented across the codebase and merging it into one. "We're on the path to fully automating the maintenance of our apps." The engineers left over get to ship products and talk to users.

Experienced engineers over-specify — and that's exactly the wrong instinct for modern models

Decades with deterministic systems built a precise, controlling instinct. With modern models, it backfires. "People give Claude way overly specific instructions — and for modern models, that's really not the way to do it." Describe the task, the guardrails, the exit criteria — then stop. The engineering brain wants to direct implementation; the model needs to be treated like a coworker. Un-learning this, Cherny says, is a journey — but it's the main thing separating people who are getting real leverage from those who are just spinning.

Opus 5 runs for days or weeks without scaffolding — loops and keep-alive hacks are already obsolete

"It can go for days, weeks, months at a time. It just won't stop. You don't even need scaffolding. You don't need slash goal." When Cherny asked the YC audience how many had run a task for more than two weeks, no hands went up. That's the gap. The model with a fortnight of autonomous capacity already exists — the only missing piece is someone willing to hand it a problem large enough, and a verification signal clear enough to keep it moving.

The suppression is lifting — and most builders haven't noticed yet

The pattern across every example Cherny gave is identical: the model already had the capability, and something — a product, a prompt, a scaffolding assumption — was suppressing it. That layer keeps thinning. Every six months, another scaffold becomes unnecessary, another category of "too hard for AI" becomes routine. The builders who compound fastest won't have the cleverest prompts. They'll be the ones who got out of the way first.

---

Topics: Claude Code, AI agents, prompt engineering, model elicitation, product overhang, agentic AI, Opus 5, dynamic workflows, test-time compute, prompt injection, system prompts, AI development, Y Combinator, LLM tooling, autonomous agents