worktree-parallellisted

Parallel sessions multiply throughput only when the tracks are genuinely independent and each track closes its own verification loop before merging. Worktrees share the repo's `.claude/` (it's in-tree), so every parallel session gets the same skills, hooks, and memory for free. ## Hold in mind 1. Tracks must be independent: no two tracks editing the same files, or the merge eats the savings. 2. One task per session: a session's context is one workspace — never multiplex tracks in one conversation; 3–5 concurrent sessions is the practical sweet spot. 3. Each track verifies green in its own worktree before any merge. 4. Worktrees are disposable; branches carry the work. Clean up after merging. ## Steps 1. Split the task into independent tracks. For each, write one paragraph: scope, files it will touch, its "done means" check. If two tracks overlap on files, merge them into one track. Gate: show the user the track table and wait for a yes before creating any worktree. 2. Launch each track in its own worktree: `claude --worktree <slug>` — Claude Code creates and manages the worktree itself (add `--tmux` to give each its own tmux session/pane). For delegated parallel edits inside one session, a subagent with worktree isolation is the equivalent. 3. Give each session its kickoff prompt: the track's scope paragraph, its "done means" check, and the instruction to run `/verify-loop` before declaring done. 4. Monitor the fleet: number your terminal tabs per track and enable system