Claude Code shifts from instructions to goals for tasks
A developer describes how their use of Claude Code and Codex has evolved from giving detailed step-by-step instructions to setting broader goals and letting AI handle execution. Instead of specifying each file, command, or folder move, they now outline the desired outcome—such as organizing side hustle materials or adding a search feature—and let the AI explore existing structures, ask…
Key points
- Claude Code users now set **goals** instead of detailed instructions, letting AI explore contexts and propose solutions
- Anthropic’s data shows humans plan **70%** of tasks while AI executes **80%**, per 400,000 session analysis
- Verification relies on **evidence** (e.g., test logs, file changes) rather than AI’s self-reports to ensure safety
Anthropic’s analysis of 400,000 Claude Code sessions shows humans typically decide 70% of planning (e.g., what to build or how to judge completion), while AI handles 80% of execution (e.g., which files to modify or commands to run). The author frames this as a division of labor: humans set goals, boundaries, and success criteria, while AI executes within those constraints. They emphasize verifying work through evidence—like Git diffs, test results, or artifact outputs—rather than trusting AI’s self-reports. OpenAI’s guidance on Codex and GPT-6 Astra aligns with this approach, stressing structured goals, constraints, and iterative checks to avoid unintended scope creep or safety risks.
Model page: GPT-6 Astra →
The story so far
2 episodes →- Claude Code shifts from instructions to goals for tasksthis story
Give AI a 'Goal' instead of 'Instructions'. How work styles have changed with Claude Code and Codex
note.com · 12 October 2026
Loading the full article…
This text was published by note.com and written by suke. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- The Roadmap to Mastering Agentic AI Systems in 2026 · 1 src
- VS Code 1.141 unifies Copilot and Codex sessions, adds agent security · 1 src
- IBM's 2016 analytics agent anticipated modern AI loops, author reflects · 1 src
- ChatGPT’s Intelligent UI turns answers into interactive apps · 1 src
- Nate Herk releases free Claude Code skill for AI advisory board · 1 src
Comments
via GitHub Discussions