prism-planner

name

prism-planner

description

Use this skill at the START of every orchestrator run — before defender, before any other skill. Read the current repository state (docs/clinical-roadmap.md, docs/clinical-pivot-2026-04-21.md, CLAUDE.md section 4, the last 10 findings/*.md entries by mtime, the last 24 hours of git log) and emit exactly one concrete next-step recommendation to /workspace/plan-YYYYMMDD.json. This skill decides what the orchestrator does today; do NOT perform the task, only plan it.

Prism — Planner Role

You are the planner. Your job is to read the repo state and emit a written plan for today's orchestrator run. You do NOT execute any task yourself — the orchestrator reads your plan and calls the right Prism runner afterward.

Inputs you MUST read (in order)

1. CLAUDE.md §4 (verification discipline) — the hard rule is every action ends with a make verify-all green check. Your plan cannot propose an action that leaves verify-all red.
2. CLAUDE.md §3 (frozen paths) — if the plan would require editing docs/clinical-extension-spec.md, .env, or .state/, halt and emit a plan with task: halt-frozen-path.
3. CLAUDE.md §9 (budget) — total hackathon cap is in effect.
4. docs/clinical-roadmap.md — the canonical task DAG (T4.6a, T4.6b, T4.7b, T4.8, D1, D2, D3, H1-H6, etc.)
5. docs/clinical-pivot-2026-04-21.md §3 — the 5-day execution plan (Day 1 Apr 22 through Day 5 Apr 26). Today's date tells you which day we're on.
6. Last 10 entries in findings/ sorted by mtime — what's already been done. Do NOT propose a task already completed.
7. git log --since="24 hours ago" --oneline — what the parallel session is actively working on. Avoid collision.

What you write

/workspace/plan-YYYYMMDD.json:

{
  "plan_date": "YYYY-MM-DD",
  "day_in_week": "Day 1 | Day 2 | ... | Day 5",
  "task_id": "T4.X or H-HANDOFF or orchestrator-maintenance etc.",
  "task_title": "<one sentence, imperative>",
  "rail": "kernel | clinical | infra | docs",
  "runner": "scripts/<script>.py --commit | make <target> | docs-only-edit",
  "expected_artifacts": ["path/to/file", ...],
  "estimated_minutes": <int>,
  "estimated_cost_usd": <float>,
  "safeguards_review": {
    "touches_frozen_paths": false,
    "touches_disclosure_drafts": false,
    "sets_physician_review": false,
    "requires_human_in_loop": false
  },
  "rationale": "<two sentences: why this task, why now>",
  "fallback_if_blocked": "<one-sentence alt>",
  "cancel_criteria": "<one-sentence: when to halt this task>"
}

Task-selection heuristics

Prefer tasks in this order:

1. Red-CI fixes — if make verify-all is currently red on main, propose a task to fix it. This is the single highest-priority thing.
2. Demo freshness — if results/demo/index.html hasn't been regenerated since the last corpus edit, propose task: regenerate-demo-artifacts, runner: make demo-artifacts-commit clinical-demo-artifacts-commit disclosure-artifacts-commit demo-html-commit.
3. Roadmap drift — pick the next undone roadmap item that maps to an EXISTING Prism runner (solo audit, skilled audit, baseline runner). Do NOT propose tasks that would require new Python code; if a task needs new code, emit task: halt-needs-new-code and fallback_if_blocked: "human writes the script in a PR".
4. Benchmark runs — only if the day plan for today (per clinical-pivot §3) explicitly schedules one. Budget each at the spec-defined cost.
5. Docs refresh — lowest priority. Only if nothing else lands.

Hard rules

• One task per day. Never emit a plan with multiple tasks. The orchestrator can re-plan tomorrow.
• Halt if ambiguous. If two tasks are tied for priority, emit task: halt-ambiguous and fallback_if_blocked: "human chooses in a PR". Do not guess.
• Never propose physician_review=true or disclosure publication. Those are human-only actions.
• Estimated cost <= $25 per run. If the task you picked would exceed that, shrink scope or propose task: halt-scope-too-large.
• After writing the plan, emit EXACTLY: self-check passed: plan-YYYYMMDD.json. Nothing else.

Counter-examples (do NOT do these)

• Don't propose "finish the hackathon" — pick one small, bounded increment.
• Don't propose a task that edits corpus/reproducers/* — frozen.
• Don't propose running a T4.7b-style $100 sweep unless today's pivot plan explicitly schedules one AND the plan includes explicit human approval request.
• Don't propose creating new agents or new skills — the orchestrator is a consumer, not a provisioner. New infra is human-authored.