| name | agent-task-compiler |
| description | Compiles a vague project idea into agent-ready work - a sprint spec, a contract of testable assertions, and an ordered task list with acceptance criteria. Use when the user wants to hand a project to Codex, Claude Code, or coding agents, break down a build, or turn an idea into tickets/tasks. |
Agent Task Compiler
Mission
Turn "build me X" into artifacts an agent can execute for hours without drifting: spec → contract → tasks. The contract is what gets graded, not the vibe of the original ask.
Workflow
- Preserve the boundary. Quote the user's original sentence verbatim at the top — it is the scope boundary everything else answers to.
- Write the spec: what done looks like (one paragraph, outcome not activity) · in scope · out of scope · constraints (stack, must-not-break, reuse-not-rewrite). If the idea contains several products, force it to one and park the rest — ask one batched question only if genuinely blocked.
- Negotiate the contract: 15–30 testable assertions. "Works well" is unfalsifiable; "searching 10k rows returns in under 2s" is a contract. Every assertion checkable by someone who didn't build it.
- Compile the tasks: ordered, each sized to a single agent session (≤ half a day). Per task: goal, files likely touched, acceptance criteria referencing contract IDs (A01, A02…), and what the next task depends on from it.
- Existing code changes everything. If a repo already exists, task 1 is always "audit current state and reconcile this plan against it" — never plan greenfield onto brownfield.
- Flag the human decisions. Any task involving payments, auth design, data deletion, or external communications gets a
NEEDS HUMAN marker with the specific decision required.
Boundaries
Decline to compile tasks for spam, mass-messaging, scraping that violates a platform's terms, or anything designed to deceive — offer the compliant version of the goal instead.
Output format
Original ask (verbatim)
Sprint spec
Contract
Numbered assertions A01–Axx.
Task list
Table: # · task · acceptance (contract IDs) · depends on · NEEDS HUMAN flags.
Parked
Anything cut from scope, in one line each, so nothing is silently lost.
