Anthropic deleted 80% of Claude Code's system prompt when Opus 5 shipped, and the model got smarter without those lines. Most of the instructions were just patching behavior the model now handles on its own. Boris Cherny, creator and head of Claude Code, explains:
> "The model reads this instruction every single time you use it"
The four parts of a working prompt
The framework he describes has four parts — here's what a prompt that actually works needs to have:
1. The task. One sentence describing the desired outcome. No roles, no "act as a senior developer."
2. Constraints. What not to touch, what boundaries to stay within, what tools to use.
3. Exit criteria. A number or checkable condition the agent can use to know for itself when it's done.
4. A way to verify. Access to a tool that measures the result without you — a profiler, tests, screenshots, whatever you'd use to check it yourself. Cherny calls verifiability the thing people miss most.
An 18-word prompt
All four parts fit in a single line. Here's Cherny's prompt, from a reply to a post by Vercel's founder on July 23:
> Hey Fable, use a dynamic workflow to get p95 under 300ms. Don't stop until you're done, use the profiler.
p95 here is the threshold that 95% of requests fall under. The task, the constraint, the exit criteria, and the verification tool — all packed into eighteen words.
There's no secret formula
This isn't just one Anthropic employee's personal opinion. In April, OpenAI's guide to GPT-5.5 advised starting the prompt from a blank page when migrating to a new model and using the shortest version that still preserves the task's meaning. Nobody has a secret formula here, Cherny says:
> "Everyone's looking for one clever trick. It doesn't exist"
What helps is clarity — prompt length has nothing to do with it.
Better to strip the prompt of numbered steps, walls of chat logs, and a bloated CLAUDE.md. Cherny recommends deleting your instructions, skills, and hooks entirely every six months, and bringing back only what the model keeps tripping over the same way, twice — the same minimalism principle we covered in You are your context.