ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Anthropic's Claude Code Team Member Addy Osmani: You May Be Over-Prompting the New Version of Claude

Addy Osmani, drawing on practices from the Claude Code team, points out that the new model requires fewer prompts than most users currently provide. He offers specific methods for streamlining: giving the model a single, clear objective along with verification conditions to work autonomously, regularly auditing old prompts using the /doctor command, and defining strict boundaries for the use of subagents.

2026-10-29 14:13 GMT+8

Addy Osmani, a member of Anthropic's Claude Code team, published a long-form post on X (formerly Twitter) with the core assessment that newer versions of the Claude model require significantly less prompting than what most users are still providing. He cited a specific figure: the Claude Code team removed over 80% of the content from the system prompts for Claude Opus 5 and Claude Fable 5, resulting in no measurable degradation in coding evaluation benchmarks.

Based on this practice, Osmani’s operational advice is to give Claude a single prompt containing a clear goal and verification conditions (i.e., how the model should confirm the task is complete), and then let it work. Start with the /effort medium setting and adjust as needed.

For situations where you are not yet clear on exactly what needs to be done, he suggests first letting Claude interview you until it can draft a specification document. Then, start a new session and use that spec to build what you want. The key mechanism here is that if a task can be verified by a pass/fail result or a numerical target to beat, and the report format is readable by the model, Claude can continue working autonomously without your supervision.

Osmani provided an actionable audit method: when a new model is released (such as Opus 5.5), remove some of the content from your own instructions and only add back those elements that the model actually misses. The /doctor prompt-audit command in Claude Code flags instructions written for older models. He recommends auditing your workflow every few months to ensure you are not writing instructions based on how older models used to behave.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

Regarding the use of subagents, Osmani’s judgment is quite clear: Claude will delegate tasks to subagents on its own, so users rarely need to request them explicitly. Only explicitly request subagents for large, independent tasks, such as long audits or migrations. For small tasks, subagents multiply costs and latency. His core view is that while subagents can parallelize work or provide a second opinion, they do not make outcomes smarter by default. For agents you build yourself, a single agent equipped with skills may outperform a subagent design in quality, often at lower cost and latency—a judgment he explicitly limits to personal observation.

It is important to note that the data regarding the 80% reduction in system prompts comes from the Claude Code team’s own internal practice reports; independent benchmark methodologies or specific evaluation sets have not been publicly disclosed. Osmani’s recommendations are largely based on his first-hand observations as a team member, representing attributable practical methods rather than peer-validated experimental conclusions.

Osmani also mentioned that Claude Code recently set auto mode as the default startup mode for interactive terminal sessions. In this mode, a classifier analyzes Claude’s behavior rather than relying on traditional permission prompts. He explicitly warned that this is not a guarantee of safe behavior—neither is blindly approving every prompt. When Claude reports completion, you should ask it to show you the commands, outputs, or screenshots.

Overall, the transferable insight from this methodology is that while model capabilities are advancing, user workflows often lag behind. Regularly auditing prompts, trusting the model’s autonomous execution under clear verification conditions, and avoiding over-engineering subagents for small tasks are practical paths to reducing usage friction and costs.

Sources

  • aihot.newsAIHOT↗
Continue researching →