Task Driver

Problem

Ambiguous complex tasks often jump straight into execution when goals, scope, acceptance criteria, or risk boundaries are unclear. This is common across multiple files, modules, external systems, or actions such as branching, merging, deletion, overwrite, and cost-affecting operations. Task Driver keeps the task in a discussable, confirmable, and verifiable state before implementation.

How It Works

The skill follows a fixed workflow rather than adding extra stages:
- Entry gate: the first reply only lists the current stage, identified items, and the first open question, without proposing code or file changes.
- Item-by-item discussion: each turn handles one decision point, separating facts, suggestions, and confirmed decisions.
- Planning and acceptance: plans are organized by confirmed requirements, with confirmed scope and implementation approach; final acceptance goals are generated only after plan confirmation, avoiding treating tests, scripts, or docs as user goals.
- Execution and self-check: when host task tools exist, tasks map to confirmed acceptance goals; each goal is checked as met, unmet, or blocked, with at most three fix attempts for repairable issues.
- Review and delivery: optional subagent review can verify evidence and regressions; unresolved issues produce a conclusion instead of being hidden as completion.

Boundaries

Use it for unclear objectives, multi-module work, or permission, release, cost, rollback, and irreversible-risk scenarios. It is not intended for simple answers, one safe command, or explicitly direct low-risk edits. A normal “confirm” does not approve the next stage, and commit, push, merge, delete, overwrite, or cost actions still require explicit user authorization.