Rewinding is cheaper than patching

22 April 2026 Michael Engel Costs Prompting

Anyone who works with an AI for a long time should tend the history like a desk. A mistake is rewound, not painted over.

In depth: Keeping context clean – fewer tokens, clearer answers

In Claude Code, every message sends the entire history so far. A failed attempt that you correct in the chat therefore stays and costs tokens again with every further message. AI LABS summarises an Anthropic post: press Esc twice or /rewind and continue writing with the new direction (according to the docs, this does not rewind changes made through Bash or subagents), /clear for a new task, pass long work on via a handoff file. All commands are in the Claude Code docs (docs checked). What remains open is whether to compact at all: in April, AI LABS advises summarising manually at a chosen point, in August, avoiding /compact altogether (28 August).

Tool and alternative

Everything is built in at no extra cost. /usage and /context show where the usage goes. According to the docs, a command-line tool is leaner than an MCP server for the same service, even though Claude Code now loads an MCP server’s tool descriptions only when needed.

Sources

AI LABS, “This Is The Anthropic Team’s Greatest Advice”, 22 April 2026 (5:52–6:25, 7:31–8:03, 8:03–9:07, 11:15–12:20)
AI LABS, “How To Never Run Out Of Codex and Claude Usage Limits”, 28 August 2026 (3:17–4:22)
Claude Code docs, Commands and Checkpointing, retrieved 11 October 2026

Advertising in the source: The 22 April video sponsored by Verdant (6:25–7:30). Both videos with self-promotion for the paid community. In the 28 August video, Supabase is praised without an ad disclosure.

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab