Graphic tutorial: Equip Claude Code with Planning with Files and from today/clear will no longer clear your tasks图文教程：给 Claude Code 装上 Planning with Files，从今天起 /clear 不再清空你的任务

Execute npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g in the terminal, -g indicates that it is a global skill and can be used for all projects.

Claude Code users can also install through the plugin marketplace; the repository docs/installation.md has corresponding methods for other hosts such as Codex, Cursor, OpenCode, etc.

Restart the Claude Code session and the hooks will take effect - you don't need to memorize any new commands after that.

Give it a job that will usually lose half the context, for example, "investigate the test coverage of this warehouse, write a phased plan for supplementary testing and implement it step by step".

Observe the agent's first reaction: it will create three files: task_plan.md (stage and check), findings.md (discovery and decision), and progress.md (operation flow) in the project root directory before starting work.

These three files are its external brain; you open them at any time and see the latest true state of the task, not the state of the agent's “verbal report.”

At halfway through the task, enter/clear to clear the context - this is the “amnesia moment” for the vast majority of proxy tools.

Then just say "continue with the task": the hook will re-inject the planned state on the disk into the context, the agent can read through the three files and then go to the current stage, no longer understand the project from scratch.

Control and feel the official benchmark: the repositioning after the reset drops from 13.3 rounds to 5.0 rounds, the longer the context and the more complex the task, the more obvious the gap.

In gated mode, agents are not allowed to announce the closing of work before each stage is completed: there are still unchecked stages in the plan, and stopping the hook will push it back to continue to do - curing the old problem of "saying it is done".

The plan file is ordinary Markdown: you can manually change the stage division and add constraints to the findings.md, and the agent will follow your version in the next round; it is recommended to submit the three files together to git, and the task history can be traced naturally.

Closing check: All tasks_plan.md are checked, and progress.md records are complete, indicating that this task is really closed loop; the next session can be quickly reviewed with/catchup.

This article is compiled from public sources and ships with the matching resource. Product features and pricing are subject to the official page. Resources are for learning and exchange only — please respect the original license.