brainstorm

Authoring a minimise plan under control

What this is

A skill one level above the CLI. You review the exact plan and approve it before anything runs — instead of hand-writing blank-page YAML or trusting a vibe-coded plan you never saw. It changes zero existing CLI behavior; it is purely additive. Delete it and mini is exactly as it was.

Prerequisite

If mini --help fails, tell the user to run /minimise:setup and stop.

Triage first — loop or job

A job runs fixed, known, finite steps once. A loop iterates plan → implement → evaluate until a goal is met. Infer which from the user's one-liner, then confirm it with them — this choice decides which schema the survey fills, so getting it wrong wastes the whole interview.

The interview

Reverse-engineer the survey from the triaged schema: for a loop, fill goal + dimensions + max_iterations + reuse; for a job, the task breakdown + gates. The exact field contract is fixed — before authoring a job plan, follow ../job/reference/plan-schema.md (estimated_duration_min is required on every task and hook, or mini job new rejects it); for a loop, the full spec is inlined in /minimise:loop. Four settled mechanics:

1. Ask cold, retrieve after (order A). Ask the problem-framing questions cold first, then retrieve from the corpus against the fuller statement. Better-conditioned retrieval, and no anchoring to the nearest neighbor before the problem is framed.
2. Every card is escapable. Each AskUserQuestion card carries concrete options plus an explicit "You suggest" pick plus free-text "Other" — delegating back to the agent is first-class, not a fallback. In the cold phase "You suggest" grounds in the one-liner; after retrieval it is corpus-grounded.
3. One concept per turn, ~3 cards. Not all-at-once, not one-field-at-a-time. Concept-grouped: (a) goal + stopping condition, (b) dimensions + max_iterations, (c) reuse (after retrieval).
4. Retrieval = a subagent's semantic judgment over mini persona list + recent loops. No keyword matching, no embeddings index. It returns a single top pick to fork and the relevant personas to reference (by versioned name, not inlined — so upgrades propagate). The subagent drafts; the main agent asks — AskUserQuestion cannot run in a subagent.