Stop Over-Prompting Claude Code: Simpler Prompts Yield Better Results

Simplify Prompts for Modern Claude Models

Addy Osmani, Anthropic Claude Code team member and former Google engineering director, argues that users often over-prompt Claude. With models like Opus 5.5, a single prompt stating the goal and how Claude should verify completion is sufficient. Start with /effort medium and adjust reasoning intensity as needed.

Use Acceptance Criteria for Autonomous Execution

Define executable acceptance criteria so Claude can self-verify. If Claude can run checks (pass/fail or numeric targets) and read reports, it can iterate independently. For long-running tasks, use /goal to set conditions evaluated by another model each round. Ethan Mollick suggests adding checkpoints for interim prototypes.

Audit and Reduce Prompt Bloat

Claude Code removed over 80% of system prompts for Opus 5 and Fable 5 without measurable benchmark regression. Osmani recommends users similarly prune accumulated prompts: delete portions, observe omissions, then restore only necessary instructions. The /doctor prompt-audit command identifies prompts written for older models.

Right-Size Sub-Agent Usage

Claude can self-dispatch work; explicitly request parallel sub-agents only for large, independent tasks like comprehensive reviews or migrations. For small tasks, multiple sub-agents increase cost and latency. Osmani observes that a single well-equipped agent often outperforms multiple custom agents at lower cost.

Leverage Cloud Sessions for Continuous Work

Use claude.ai/code, the Claude app, or claude --cloud to run sessions in isolated VMs that clone the repository. Push local commits first. Monitor progress via phone, answer questions, then review diffs and create PRs. Reserve local execution for tasks needing machine-specific resources.

Manage Multiple Sessions with Projects

Claude Code's Projects feature (in public beta for Pro/Max users) organizes long-running conversations. Each task gets a separate thread, often running in parallel cloud sessions. Project requirements propagate to new threads; the Overview page tracks progress, reviewable PRs, and pending user input.

Verify Results and Automate Code Review

Auto mode is now default for interactive terminals, but Osmani warns of limitations. When Claude reports completion, ask it to show executed commands, outputs, or screenshots. For code review, let Claude run a first pass: the Claude Code team uses Claude to review and fix most changes before human review. Use Team/Enterprise Code Review or lightweight /code-review for a fresh sub-agent to examine diffs, focusing on requirements over style.

Regularly Reassess Workflow

Every few months, audit which prompts can be trimmed, which tasks can be fully delegated, and which results still need human verification.