Make your AI coding agent's handoff checkable: point every "done" at something real

When a session of the Claude Code AI runs out of context or you pause for the day, the following session begins with the previous one's writing. However, the AI most likely made an error in that summary. To prevent issues, a habit must be developed. Each claim should have a pointer, not just the word "done." These pointers should be linked to something verifiable, such as a commit hash, a file within that commit, or a line in a log.

If the AI cannot provide such a pointer, the statement should be marked as "not done." Before the new session starts working, the pointers should be checked using a few lines of shell script. The script verifies if each commit or file exists. The count of existing and non-existing items is then displayed to the user before the new session begins.

It is crucial to understand that a pointer existing does not guarantee the truthfulness of the statement. The check merely indicates which statements are based on actual work and which are not. This information can help identify claims that rest on nothing and guide the investigation to the relevant issues. The checkpoints should be kept in the project's repository in plain text Markdown and JSON files.

This way, any agent or tool can easily read and use the information, and it will travel with the code. The Succession tool, which automates this process, is available for free and is licensed under the MIT license. It is a small plugin that integrates seamlessly with Git and works on any POSIX shell.