What problem does it solve? Non-trivial tasks fail when agents act on vague goals, wrong assumptions about the system, or missing review. This Skill structures work into three phases — Calibrate, Align, Plan — with evaluation woven through every step, so large or ambiguous tasks stay grounded in reality and verified before delivery. ## Core Features & Use Cases - Calibrate: Sync agent and human mental models with a codebase, system, or domain through an interactive exploration loop that tracks drift and produces a calibration record. - Align: Grill the user with decision-tree question rounds, each carrying a recommended answer, until an alignment document with concrete success criteria is confirmed. - Plan: Decompose large or parallel work into an executable task graph with dependencies, checkpoints, risks, and capability matching. - Work & Evaluate: Execute step by step with adversarial self-challenge, routing failures back to the owning phase, and gating large tasks behind an independent adversarial review by a fresh agent. - Use Case: A user asks to migrate a production service. The agent calibrates the current architecture, aligns on success criteria and constraints, plans the parallel migration steps, executes with continuous self-verification, and spawns an independent reviewer before delivery. ## Quick Start Use the ava-workflow skill to align on the goal, plan the steps, and execute this task with continuous evaluation and an independent review before delivery.
