The context · September 2025
Anthropic’s September 2025 Sonnet 4.5 announcement included updates to Claude Code, with checkpoints among the features supporting coding workflows. [1]
Confidence needs a recovery path
A developer can explore a larger change more comfortably when returning to a known working state is straightforward. The same applies to assistant-led work. Recovery should preserve the user’s existing changes and distinguish a deliberate experiment from the current accepted implementation.
A checkpoint needs meaning
Saving state after every action creates a history, but not necessarily an understandable one. Useful checkpoints correspond to decisions: an investigation completed, a failing case reproduced or a behaviour corrected. Pair the saved state with a short explanation of what was learned and what remains uncertain.
Review in coherent pieces
Small, meaningful changes make it easier to spot an accidental behaviour change or an unnecessary abstraction. Checkpoints support this practice; they do not replace version control, review or suitable checks. The human team should remain able to explain the final code without reconstructing an entire assistant conversation.
Source & context
Anthropic · Claude Sonnet 4.5, 29 September 2025This retrospective was written for the archive in September 2026. The linked primary source documents the announcement or event; the practical interpretation and proposed approach are Sansa’s editorial perspective. Public examples do not imply a client relationship. Product capabilities and guidance may have changed since the period discussed.
Another perspective · September 2025
A review queue for assistant-generated changes
Continue readingWorking through a similar question?
Talk it through with Sansa