task-intake-contract

name	task-intake-contract
description	Build a compact, explicit intake contract for messy requests before planning or delegation. Use inside ultracode-high and other high-rigor workflows to identify the goal, artifact, constraints, acceptance criteria, missing decisions, risk level, and verification target.
model	claude-opus-4-8
effort	high

Task Intake Contract

Convert the request into a contract another agent could execute without guessing.

Fleet context

This skill is part of ultracode-high. For the full orchestration map, tool catalog, agent roster, and card schemas, read ../ultracode-high/references/runtime-context.md when running inside this repo or the installed ultracode-high skill package. Your output becomes the goal-contract that every downstream agent re-reads to prevent constraint decay.

Fields

goal:                    # objective, quoted verbatim where load-bearing
artifact:
audience:
decision_supported:
constraints:
acceptance_criteria:
scope_in:
non_goals:
known_inputs:
unknowns:                # each: { q, disposition: search | compute | ask | flag }
missing_decisions:
decisions:               # appended as the run proceeds: { decision, rationale }
risk_of_being_wrong: low | moderate | high
verification_target:
permission_boundaries:

Persistence

Write the contract to .claude/ultracode-high/<run-id>/CONTRACT.md at intake so it survives main-context compaction. Re-read it every iteration and inject it into every subagent's work package; workers should receive the contract plus their one item, never raw conversation history, so they cannot drift. Append decisions with rationale as the run proceeds and never silently drop a constraint.

Rules

• Infer reasonable defaults only when the decision would not materially change.
• Ask or return needs-input when a missing decision changes the deliverable.
• Quote user constraints verbatim when they are load-bearing.
• Separate facts, assumptions, and unknowns.
• Give every Unknown a disposition: search (find it), compute (derive it), ask (block on the user), or flag (surface it as a residual risk). A definite claim for an unresolved Unknown must fail verification rather than be invented.
• Include the smallest next artifact that would unlock progress when blocked.