Boris Cherny: We Cut 80% of Claude Code’s Prompt

Predictions and Expectations:

• Boris anticipates that within the next two to three years, models will exhibit capabilities currently unrealized, describing a "huge amount" of "product overhang" that startups have yet to capture.
• He expects that the current "prompt engineering" trend will evolve, with skills shifting away from specific instruction writing toward "empirical" task definition and verification.
• Boris believes that "coding is solved" for the types of engineering he performs, though he notes exceptions remain for deep systems code, distributed systems, and pixel-perfect UI verification.
• He predicts that within six months, developers will find that deleting their existing "bottom D, skills, and hooks" reveals that newer models no longer require those specific instructions.
• He expects that users will frequently encounter "surprises" where models perform tasks beyond their expected capabilities if given the right task and verification tools.
• He forecasts that dynamic workflows will allow for the productive orchestration of "tens of thousands" of agents to solve complex problems.
• He predicts a future where "hundreds of agents" run daily to fully automate application maintenance, including cleanup, testing, and abstraction unification.

Timelines and Milestones:

• Boris states the shift away from "single line autocomplete" toward writing entire files happened roughly "one and a half to two years ago."
• He mentions that Opus 5 was released "yesterday" relative to the conversation.
• He notes that the capability to rewrite code bases across languages began appearing with "Opus 4" (referred to as "Fable" in the transcript context, likely a slip for a version between 4 and 5) and is fully realized in Opus 5.
• He describes a current internal experiment rewriting an Electron app in Swift that has been running for "a little over two weeks" (14-15 days) and expects it to complete eventually.
• He predicts that "every day" Anthropic will run "20 or 30" distinct routines to maintain their codebases, moving toward full automation.
• He expects that model evals will typically survive "one, two, or three model generations" before saturating and needing replacement.
• He notes that the current product release cycle involves deleting and rebuilding system prompts "every six months" or "every time a new model comes out."

Technology and Product Direction:

• Boris plans to "press delete" on system prompts and scaffolding for every new model release to rely on the model's native intelligence.
• He expects future product development to focus on "unhobbling" models, giving them harder tasks and verification tools rather than rigid, step-by-step instructions.
• He intends to use "dynamic workflows" to orchestrate complex, multi-stage tasks involving "dozens or hundreds of agents" rather than single-pass scripts.
• He plans to implement "loops and routines" (local and cloud-based) to automate repetitive maintenance tasks like cleaning dead code, shipping experiments, and unifying abstractions.
• He expects the industry to move toward "test time compute" scaling, where the value comes from the amount of tokens generated and the orchestration of agents during the task rather than just model size.
• He predicts that "product overhang" will drive new product creation as founders identify capabilities models already possess but lack tools to express.
• He expects that future users will be "empirical" scientists, observing model failures to iteratively build harnesses rather than relying on upfront theoretical design.

Market and Industry Outlook:

• Boris predicts that the "prompt engineer" job title will fade, replaced by skills focused on "context engineering" and "elicitation."
• He believes that "every single time" a new model is released, the existing system prompts and tools often become obsolete and require significant rework.
• He expects that the gap between human-written code and agent-written code will close, with "more and more kinds of code" becoming "solved" by agents.
• He predicts that the best builders will be those who can "unlearn" traditional computer science rigidities and treat models as organic "coworkers" rather than deterministic tools.
• He foresees a market where "dozens" of commercial opportunities exist for capabilities that are currently "unrealized" or undiscovered by the industry.

Company Plans:

• Anthropic plans to continue deleting "over 80% of the system prompt" with each major model release to reduce friction.
• The company intends to use "dynamic workflows" and "routines" to maintain their own internal software stack, including the CLI, iOS, Android, and desktop apps.
• They plan to "unship tools" and delete code in the harness frequently to identify the true value of each component through ablation studies.
• They intend to release "Max 20X" credits to the audience to enable long-running, large-scale experiments.
• The team plans to keep appending to their "evals" until they saturate, then discard them to create new evaluations based on empirical findings.

Financial Guidance:

• No specific financial revenue, profit, or cost guidance statements were found in the transcript.

Risks and Caveats:

• Boris warns that while coding is "solved" for his use cases, "Opus 5" is still not perfect for deep systems code, distributed systems, and "pixel-perfect" UI verification.
• He notes that models may "stumble over architecture" or other specific elements, requiring iterative fixing of the harness.
• He cautions against "over-specifying" tasks, as models may not follow rigid, step-by-step instructions as they did in the past.
• He mentions that "system prompts" can get in the way if they are not carefully managed or if they correct for behaviors the new model no longer exhibits.
• He notes that "evals" have a short lifespan and may need to be discarded after "one, two, or three model generations" due to rapid model improvement.
• He warns that relying on "LinkedIn influencers" or "Twitter" for "one weird tricks" is a failure mode; the correct approach is empirical testing.

Confidence and Disagreement:

• Boris expresses high confidence ("100%") that every developer should "press delete" on their system prompts and skills to see what the new model can do.
• He states with certainty that "Opus 5" does not seem to be prompt injectable anymore, describing it as a "new frontier" compared to previous versions.
• He is "not sure" of the exact number of agents spawned in the two-week Swift rewrite task, estimating "thousands, tens of thousands" but noting he needs to check.
• He qualifies his claim that "coding is solved" by explicitly limiting it to the specific types of coding he does, acknowledging areas where the model still struggles.
• He believes with certainty that the best approach to building agentic products is "empirical" and "scientific," rejecting theoretical upfront design.