Anthropic has stripped more than 80% of the system prompt that powers Claude Code, the company’s AI coding tool, arguing that its newest models need less hand-holding than earlier ones did.
In a July 24 engineering post, Anthropic staffer Thariq Shihipar wrote that the instructions guiding Claude Opus 5 and Claude Fable 5 inside Claude Code were cut from roughly 800 tokens to about 164, with no measurable drop on the company’s internal coding evaluations.
From rules to judgment
The post frames the cut as part of a broader shift in how Anthropic approaches context engineering — the practice of designing everything a model reads before it answers, including system prompts, tool descriptions, and memory files.
Older prompts spelled out narrow rules, such as telling the model to never write multi-paragraph docstrings. The rewritten version instead asks Claude to write code that reads like the surrounding code, trusting the model’s judgment rather than boxing it in with examples. Anthropic says newer models can find such examples constraining, since they pull output toward the sample rather than the underlying intent.
Shihipar described five other changes alongside the shift from rules to judgment: dropping tool-use examples in favor of clearer tool interfaces; moving detailed guidance, like code-review instructions, into on-demand skills instead of loading it upfront; removing instructions that were repeated in both the system prompt and tool descriptions; replacing manually maintained memory files with automatic, contextually triggered memory capture; and leaning on richer references, such as test suites and real code, instead of prose specifications.
Why it matters
For developers building on Claude Code or their own Claude-based agents, the change signals that heavily scripted system prompts may now do more harm than good with the latest model generation. Anthropic frames lighter prompts as cheaper and faster to run, since fewer tokens are processed on every turn, though the company has not published independent benchmark comparisons beyond its own internal evaluations.