Write a plan

One document holds the design and the plan. What the interview settles is written once, there, and each task points into it, narrowed to its scope.

The plan is written for an agent that was not in the conversation. Read alone, it gives every decision, its reason and every detail the work needs, so that agent is as aligned as the person and you are now. Nothing settled lives only in the chat.

No code in the repository for the thing being planned until the plan is approved.

1. Open YYYY-MM-DD-<topic>.plan.md before the first question, through the aiview skill: kind plan, tags = topic, in the group aiview gives a plan, shared with its task documents when a task needs one. With a roadmap, the slot's slug is a tag. Its two parts, Design and Execution, are references/skeleton.md, read before writing the first section. A board on the topic is the starting point: its agreed decisions become rows, numbered in this plan with the board's id in their source (board <topic>.board#D2), and are not asked again.
2. Run the interview skill on the decisions table, and write each section as its decisions land. If the person declines it, offer its short mode; a decision still unasked is a row with source agent default. A request that is several independent systems is split first, one plan each. One plan per increment: it stops where the next increment would begin.
3. Every major area is a deep module, designed as the deepen-module skill says, in the plan's modules section. A module the interview cannot settle there gets its own module document, and the plan quotes its entry. Each guide that applies (the shipped guide-* skills, the project's .claude/skills/guide-*) gets its decision for this plan in the plan's Guides section.
4. When no row is open, ask once what else this should cover.
5. Settle the phases before the cut: the plan runs on N seats, phase by phase, on the board of the swarm-board skill. Read references/phases.md now; then ask phases:, seats: and verify:, one per message, as the orchestrate-plan skill's references/mandate.md says.
6. Cut the phases as references/phases.md and the orchestrate skill's references/tasks.md say, and draw the tracker to references/tracker.md, from references/tracker-skeleton.md, both read first. The execute-plan skill keeps it from then on. aiview tracker check <plan> and scripts/decisions.mjs <the plan's folder> after every edit: the last says whether a topic, an id or a citation collides or points nowhere.
7. Before the mandate, dispatch a fresh subagent on the cold-read skill, with the plan's id and none of the conversation; discuss with it as a peer, explaining your intent and asking how it read each part, until you understand each other; every point you settle goes into the plan, and it re-reads until it says ready.
8. Close the interview with the orchestrator's mandate: the questions of the orchestrate-plan skill's references/mandate.md, the answers in the plan's Mandate section and in the one line under the tracker. tests: at the plan end adds the tests phase to the tracker then. Whatever the mandate changed in the plan, the cold reader re-reads, until it says ready again.
9. Stop and ask the person to approve the plan. The request opens on the agent default rows, numbered, contract ones first, "4 decisions are mine: … confirm them, or name the ones to change", and the link to the plan comes after the list. The approval goes in the status line with the defaults it confirmed, then aiview plan check <plan>. No task is written now: each is written at its phase's open.

Red flags