Lets Claude make several edits and their check, or a language-server lookup or rename, in one step through the real tools; guidance on chaining, which a subagent's task carries too; and a short note when a step cost a model request of its own for nothing.
A hand_off tool: Claude gives a self-contained task to a cheaper worker model (Haiku for lookups and mechanical changes, Sonnet for build work) that stays warm between tasks, with where the change reaches from the language server, the diff and a check run by the tool as evidence, a review by the main model, and a log of what it saved and how often the work came back.
Writing to a length in one draft and cuts, not rewrites. When your prompt sets a word limit, it carries your setting to plan each part's share and land the length in one draft, or one more write; each write of a document then shows its words as wc -w counts them, and when it is well over, each section's target. Silent without a limit. Nothing leaves your machine.
Shows long-running and background jobs with a percentage and an estimated finish time, read live from their output.
The claude-mods modpack: statusline, smart-compact, delegate, cells, eta, seek and draft, installed together and built to work together, and route, an optional mod that needs an outside classifier, off until you turn it on.
Optional, and the one mod that needs an outside classifier model: hands feature and test work to delegate's cheaper worker. A fast classifier (TypeSafe's Jev or ScaleDown, with your API key) reads your prompt, and only where it is feature or test work does Claude get the order to implement it through a hand-off. Off until you turn it on and give it a key for one of them.
Has Claude look around your code in few, lean steps. Every request re-reads the whole conversation, and whatever Claude reads stays in it for every later request; so your prompt, and the task of a subagent Claude starts, carries your setting to batch the searches and reads Claude can already name, to read definitions or line ranges instead of whole files, and not to read again what the session...
Keeps an idle session's prompt cache warm, or compacts it just before it expires, whichever is cheaper, and says on the status line whether the next prompt is ready or cold.
One line under the prompt: the model and its effort, how full the context is, the plan's 5-hour and 7-day use with their resets, and what other mods put on it (smart-compact's ready or cold).