What they said

In a fireside chat transcript published by Simon Willison, Anthropic’s Cat Wu and Thariq Shihipar described why their team simplified prompts for its newest models.

Shihipar explained the reasoning behind cutting Claude Code’s system prompt by roughly 80 percent. The old prompt was “over-constraining Claude” with example outputs and “do not” instructions that often conflicted with what a specific user actually needed. Removing them, according to the transcript, let newer models like Opus 4.8 and Fable 5 produce results the team called “more creative.”

Shihipar described a related shift in how he approaches the work itself: “For me, it’s that rewrites are now good.” He also talked about offsetting what he called coding-agent-induced “Deep Blue,” a reference to the chess computer that beat Garry Kasparov in 1997 and left many players feeling their own judgment mattered less. Shihipar’s answer: “The way you offset that is by being more ambitious.”

Who they are and why the take carries weight

Wu is Head of Product for Claude Code (and Cowork) at Anthropic. Shihipar is an engineer on Anthropic’s Claude Code team. Both work on Claude Code and describe the team’s experience with its prompts and workflows.

That visibility is also a limit. Anthropic builds and owns the models behind these findings, so treat them as company-reported observations, not independent tests.

What they get right, and where it’s incomplete

Wu and Shihipar agree on a common thread: as models get more capable, the old defensive habits, verbose prompts, heavy examples, and cautious scope, can cost more than they save. The team’s prompt-simplification finding and Shihipar’s push toward bigger, more ambitious tasks are two sides of the same shift toward trusting the model more.

Shihipar said the prompt changes took multiple evaluations to develop. Wu separately described external and internal evaluation suites and said their coverage was incomplete. The reduction applies to the newest models; older models kept their full prompts.

It is also narrow. The 80 percent reduction happened at a company that trains the models it’s prompting, on models tuned specifically for coding tasks. A shorter, example-free prompt that works well for Opus 4.8 or Fable 5 may not transfer cleanly to a different model family, or to a task far outside coding.

The Claude Tag detail underscores this. Wu said “Claude Tag currently lands 65% of our product eng PRs,” referring to Claude’s Slack-based feature for turning a mention into an agent task. That is a strong number, but it describes one internal team’s workflow, not a general benchmark for how often any team should expect an agent to land a finished, mergeable pull request.

Shihipar’s ambition argument is harder to test outright. Being more ambitious with agent-assisted work is reasonable advice for someone who spends all day building and evaluating these tools. It says less about whether a team with different risk tolerance, codebase, or review process should take on that much more scope at once.

Why it’s notable

Wu and Shihipar describe a move away from examples and repeated constraints inside the team that ships Claude Code, and tie it to a specific, measurable change (an 80 percent shorter system prompt) rather than a vague call to “trust the model.”

Boris Cherny argues that teams should encode more knowledge into files and rules for agents to follow. Wu and Shihipar are pointing toward the opposite pole: fewer standing instructions, more reliance on the model’s own judgment in the moment. Both can be true at once. Durable facts about a codebase belong in documentation. Stylistic hedges and example outputs may not.

What it means for builders

Do not delete a carefully tuned system prompt because Anthropic’s team cut theirs. Test it. Try removing an example of a desired response or a “don’t do X” instruction on a task where you already have a baseline, and compare the results on your current model version.

Treat prompt length as a variable to measure, not a proxy for quality. A shorter prompt can outperform a longer one, but the result depends on the model and task. Retest after a model upgrade instead of assuming old prompting habits still hold.

Shihipar’s “more ambitious” framing is worth borrowing in a narrower form: as an agent gets more reliable on a class of task, expand the scope of what you delegate to it, but do so deliberately and where your own review process can still catch mistakes. Confidence earned on small tasks should be spent gradually, not all at once.

End of article