skills/execute-zzzops/references/CREATE.md

# Create and triage

1. Use the complete BACKENDS portfolio and project evidence for duplicates/facts. Capture outcome, value, constraints, owner, dates, evidence, and non-goals; refine rather than duplicate.
2. Create one provisional human-first canonical issue through `BACKENDS.md`; use reviewed PROJECT values unless evidence requires a reviewed change.
3. Investigate outcome/proof/scope/decisions/dependencies/risks/authority and goal shape. Persist evidenced rigor-risk inputs/authorized override and require a valid effective level; ambiguous metadata blocks preparation. When unfamiliarity or decision risk could alter architecture, scope, acceptance, or quality, run a bounded blind-spot pass for known unknowns, tacit criteria, and blind spots. Use the cheapest useful evidence or disposable prototype; persist findings, not logs.
4. For consequential ambiguity apply `BLOCKERS.md`: record recommendation and continuation boundary, expose the human queue, and continue independent work without asking live. Assume only a [[bounded commitment]](../../../concepts/bounded-commitment.md) supported by reviewed policy; never invent foundational choices.
5. Write explicit scope and verifiable criteria.
6. Map parent/dependencies; use a child only for distinct verifiable, prioritized/blocked/claimed work or risk, otherwise a checklist. Obey PROJECT depth, derive inverse edges, reject cycles.
7. Derive priority/value/difficulty/confidence/owner/dates from PROJECT, `GOAL_SYSTEM.md`, and analogous evidence. Name the KPI/criterion/beneficiary; unsupported value stays low-confidence and blocked. Preserve user budgets; otherwise bound checkpoints, not cost.
8. Give each leaf an action/stop, baseline, falsifiable hypothesis, observation, smallest probe, expected signal, wider checks, and resources. Foreground change-sensitive decisions such as data models, public interfaces, architecture, and user-visible behavior; subordinate trusted mechanics. Mark generated/external/branch resources exclusive and binary paths explicit; ordinary text/integration stays advisory unless policy differs. Apply `EXECUTION_STRATEGY.md` artifact rules. If unobservable, scope the smallest harness/debug adapter/MCP server or block.
9. Set `ready` for writable/no gate; `triaged` for preparation/children/dependency; `blocked` only with no useful work; `cancelled` with rationale/replacement.
10. Update canonical state, relations, human queue, and history consistently; then continue to `EXECUTE.md` when requested.

For independent child breakdowns apply PROJECT parallel settings through `EXECUTION_STRATEGY.md`. Read-only proposals obey the lower cap; the coordinator resolves overlap/cycles and owns writes unless policy permits isolated worktrees.