| name | decisions-frontier-prep |
| description | Use when preparing hardest high-leverage work for a stronger or scarce model, building a ranked execution queue, asking for fable prep, frontier prep, round-collapse, wave queue, or model-upgrade readiness across code, design, copy, marketing, or architecture work. |
Decisions Frontier Prep
Use this to turn broad, difficult work into a small set of execution-ready packets for the best available model. The core rule is selection: keep only work where a stronger model plausibly collapses multiple iterate/fix rounds or raises the quality ceiling.
Workflow
- Detect the target harness, project, active board, and candidate repos. Use Decisions project context first; fall back to local git repos only when no board/project context exists.
- Run a broad diagnosis by domain: security, correctness, performance, architecture, design, copy, marketing, and tech debt.
- Reject phantom work. Every finding must be verified against live files, current tickets, or an explicit source artifact before it becomes a packet.
- Write packets using
references/queue-schema.md. Each packet needs concrete files or artifacts, an operation, acceptance criteria, and a machine check.
- Score with
scripts/score_frontier_queue.py or the scoring rules in references/scoring.md.
- Group packets into waves. Wave 1 should be high score and low dependency so it can prove the queue quickly.
- For DecisionsAI workflows, attach the queue artifact path to the workflow result and keep execution gated by the board/workflow owner.
Harness Rules
- Read
references/harness-adapters.md before emitting harness-specific instructions.
- Do not assume a slash command, a single CLI, or a specific model provider.
- Use Decisions workflow
pre_chain skills for execution context. Do not silently add this skill to default pre_chain.
- Keep AgentManager-style multi-agent spawning out of scope; Decisions keeps one operating agent per board.
Safety
- Archive instead of deleting.
- Never loosen sandbox, permission, branch, or deployment rules to make a packet easier.
- Do not queue work whose machine check is only "agent says done."
- Route work that a normal model can one-shot to ordinary board tickets, not the frontier queue.
