superplanlisted

> Ship signal, not noise. Parallel perspectives, one coherent plan, an approval gate that matches > the posture — human when a human is driving, self-served and recorded when none is. # Superplan Skill ## When to use - Any implementation task where the scope, architecture, or approach is non-obvious - Before writing code for a feature that touches multiple files or systems - When a written plan doc is required for human approval ## Phase 1 — Fan-out (parallel agents) Announce, as a fragment: `superplan — fan-out: planner, architect, adversary`. Odin voice does not lapse inside a skill (ADR-0011); a quotable first-person sentence in skill text is the shape that causes the lapse. Dispatch all three agents **simultaneously** using the Agent tool. **Fill `<CONTEXT>` with ≤10 paths, one line each, path + role — never file contents.** Each agent starts its own session, so `<CONTEXT>` is paid three times — on top of `CLAUDE.md` and the always-on rules, which every one of those sessions loads in full before it reads a word of the task. Two of the three copies are pure duplication: the first is the cost of doing the work at all. Run `bash .claude/scripts/context-budget.sh report` for today's figures. They are deliberately not written here — this line said `~52 KB` while the rules index said `38.8 KB` and a skill sample said `34.8 KB`, none of them agreeing and none of them produced by anything (harness:RM-0164). Every agent has its own Read and Grep; what it needs from you is