Kickoff Project

kickoff-project

A project-kickoff skill that solves "I have an idea but don't know where to start, and handing it straight to an AI coder goes off the rails."

It structures project kickoff into four phases, each confirmed with the user before moving to the next:

1. Requirements

Interviews from the user's perspective: what, for whom, core flow, boundaries. It digs before producing — when the user says "I want X", it first asks why and how they solve it today. Outputs a requirements doc that serves as the core context for all subsequent phases.

2. Architecture

Discusses tech choices and overall structure based on the requirements. Choices are driven by requirements, not tech preference, with trade-offs explained rather than conclusions alone; considers the user's existing experience and infrastructure; good-enough is enough, no over-engineering for hypothetical future scale.

3. Phased Dev Plan

Decomposes the architecture into sequential phases a coding agent can execute. The granularity anchor is a core design constraint: each phase must be independently completable by a 30B-param / 200K-context model — single concern, limited files, explicit inputs and outputs, independently verifiable completion criteria, explicit inter-phase dependencies. High-risk items get an upfront validation spike before scheduling; phase 1 is the MVP minimal validation path.

4. Project File Init

Writes all outputs into the target project so a coding agent works out of the box:

• .agents/skills/ — requirements, architecture, and dev plan each as a standalone self-contained skill, loaded by the agent on demand
• project-plan.md — dev plan + per-phase status (living doc, updated as development progresses)
• AGENTS.md — project context overview and rules

Design Traits

• One question per turn (options + recommendation + allow custom) — no question dumps, no infinite questioning
• Never asks what it can look up in the code/project
• Not sycophantic: contradictory or infeasible answers are flagged and confirmed, never silently followed
• Uncertain items are marked [TBD]; requirements the user never expressed are never fabricated
• Key decisions require user confirmation — the skill never decides for the user

The output is not documentation for humans but project infrastructure a coding agent can consume directly. Ideal for starting new projects, structuring vague ideas, or standardizing AI-assisted development workflows.