Claude Code team member on how to reduce over-prompting Claude

Model capabilities are changing, and the way we use them needs to be updated accordingly.

Recently, Addy Osmani, a member of the Anthropic Claude Code team and former engineering director at Google, published a long article about a habit many users may share: writing too many prompts and instructions for Claude.

The article's title makes the point right away: you have most likely been "over-prompting" Claude all along. As a new generation of models becomes more capable, the old approach of explaining everything and guiding step by step is also due for readjustment.

01

When using Opus 5.5, Osmani suggests first trying a simple approach: use a single prompt to clearly state what you want to accomplish and how Claude should check whether it has completed the task, then let it go and do the work. Reasoning effort can start at /effort medium and be adjusted based on actual conditions.

If users have not yet thought through their requirements clearly, they can first have Claude keep asking questions until it has enough information to write a requirements specification, then start a new session and work according to that specification.

Among these, acceptance criteria deserve extra thought. Claude believes that after a task ends, it will stop on its own; if it can run checks, use "pass or fail" to judge results, or verify against some clear numerical target, and can read the check reports, it can continue revising and checking based on the results.

With executable acceptance criteria, the model can advance tasks more independently, and users can reduce mid-course intervention. For tasks that need to run in the background for a long time, users can also use /goal to set conditions, with another model checking after each round whether they have been met.

As for when users should intervene during the process, Ethan Mollick, an associate professor of management at Wharton, added a suggestion: set staged checkpoints during the task and have the model report preliminary ideas or show a prototype.

02

So how much can prompts be streamlined?

In the article, Osmani mentions that for models such as Claude Opus 5 and Claude Fable 5, Claude Code once deleted more than 80% of its system prompts, with no measurable decline in coding benchmark performance. He also suggests that after a new model is released, users run a similar experiment with their own instructions: first delete part of them, see where the model misses requirements, then add the relevant instructions back. The /doctor prompt-audit command can help identify prompts written for older models.

In short, prompts also need regular cleanup, and rules accumulated over time can be re-validated.

Similar adjustments apply to subagents. Claude already assigns work on its own, and usually you can let it arrange things itself. For tasks with a large workload that can be carried out independently, such as comprehensive reviews or code migrations, explicitly ask it to split the work for parallel processing. For small tasks, invoking multiple subagents increases cost and waiting time, so it still depends on whether the task itself is worth splitting.

In Osmani's view, subagents are suitable for parallel task processing and providing second opinions, and the number used needs to match the scale of the task. According to his observations, when users build agents themselves, a single agent equipped with the right skills may achieve better results, at lower cost and with faster responses.

If you want tasks to continue after you close your laptop, Osmani recommends using Claude cloud sessions. Users can start them from claude.ai/code, the Claude app, or via claude --cloud. Cloud sessions re-clone the code repository in an independent virtual machine, so before starting you need to push local commits to the remote. After that, you can check progress on your phone, answer questions raised in the session, and finally review the code diff and create a PR. Tasks that require resources specific to the local machine should continue to run locally.

When running multiple sessions simultaneously, management work also increases: what each session is doing, which ones are done, and which are still waiting for replies all need to be tracked. Therefore, Osmani suggests first enabling a single session to complete tasks independently before considering running multiple sessions at once. If every piece of work requires frequent intervention, opening more sessions will also increase the time users spend switching back and forth and coordinating.

Claude Code's Projects feature takes on part of this management work. A project is equivalent to a long-running conversation; users can post emerging issues and tasks into it, and Claude will create a separate thread for each piece of work, usually processed in parallel through cloud sessions. Project requirements that users have specified will apply to subsequent new threads; the Overview page centrally displays task progress, PRs ready for review, and items waiting for user responses. According to the article, Projects was in public beta for Pro and Max users at the time.

As more execution work is handed to the model, users need to focus on verifying results. Recent versions of Claude Code have made auto mode the default startup method for interactive terminal sessions, with a classifier evaluating Claude's operations. Osmani reminds that auto mode still has limitations, and approving permission prompts one by one also requires careful judgment. His advice is that when Claude reports completion, have it show the commands and output it executed, or provide screenshots, to make it easier to verify the actual results.

Code review can also first have Claude run a round. Osmani reveals that on the Claude Code team, most changes are first reviewed and corrected by Claude at least once before being handed to humans. Users can use Code Review in the Team and Enterprise plans, or use the lighter-weight /code-review to have a subagent that has not seen the current conversation check the code diff. Starting another session and providing the requirements specification can likewise yield an independent review; during review, you can ask it to focus on requirements and functionality, with code style taking a secondary position.

At the end of the article, Osmani suggests users re-examine their way of working every few months: which prompts can already be streamlined, which tasks can be handed to the model to complete independently, and which results still need human verification.

This article comes from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: someone who follows large models, published by 36Kr with authorization.