Claude Code Todo Tool¶
The internal TodoWrite tool — a structured-but-not-structurally-enforced planning scratchpad that claude-code injects into its own context.
Mechanism (per jared-zoneraich)¶
- On first user turn, the model emits a to-do block via a tool call. Schema has a version, per-item
id(hash),title(human-readable),status, and an optionalevidenceblob. - Rules ("one task at a time", "mark completed", "keep working on in-progress", "break into sub-tasks") are purely prompt-based — not enforced in code.
- Relies on the fact that 2025-era models are finally good enough at instruction following that prompt-level structure holds. "This would not have worked a year ago."
Why it matters¶
- Forces a plan before tool thrash.
- Resumability after crashes / compaction.
- UX signal — user sees progress instead of a silent 40-minute loop.
- Steerability — user can edit the list.
Zoneraich flags it as a candidate first-class paradigm (alongside claude-code-skills) that other agent builders should copy: model-driven planning via an in-context structured artifact, not an external planner node.