Hacker News Pushes Get Shit Done as Coding-Agent Users Move From Prompts to Spec Systems

Why this repo resonated on Hacker News

On March 17, 2026, a Hacker News post linking to Get Shit Done reached 404 points and 223 comments. The repository is not selling a new foundation model. It is trying to solve the operational mess that appears after teams already have access to coding agents: prompt drift, forgotten decisions, shallow planning, and degraded output as long sessions fill the context window. The README makes that framing explicit by calling the project a meta-prompting, context-engineering, and spec-driven development system.

What the project actually ships is a workflow layer across multiple runtimes. The installer supports Claude Code, OpenCode, Gemini CLI, Codex, Copilot, and Antigravity, with Codex specifically using skills rather than custom prompts. The basic flow starts with npx get-shit-done-cc@latest, then moves into commands such as /gsd:new-project, /gsd:discuss-phase 1, and /gsd:plan-phase 1. According to the README, that process creates files such as PROJECT.md, REQUIREMENTS.md, ROADMAP.md, and STATE.md, plus research artifacts.

From “prompting” to operating system

The larger signal here is that coding-agent users are moving away from single prompt tricks and toward reusable operating systems. Get Shit Done claims to hide the complexity behind a small set of commands, but under the surface it says it is managing context engineering, XML formatting, state management, and subagent orchestration. Whether or not one adopts this specific project, the direction matters. The value is increasingly in how work is staged, remembered, researched, and handed off, not only in what one asks the model in a single turn.

That is also why HN paid attention. Many developers are discovering that agent quality degrades when there is no persistent structure around planning and verification. A system that converts a vague product idea into scoped phases, discussion context, research, and implementation state is appealing because it tries to make agent work repeatable rather than improvisational.

The tradeoff hidden in the pitch

The README also exposes the core tension in this category. The project recommends running Claude Code with --dangerously-skip-permissions for lower friction. That advice fits the product goal of uninterrupted automation, but it also shows the cost of these higher-autonomy workflows: the more seamless the system becomes, the more carefully users need to think about local permissions, tool scope, and blast radius. In other words, Get Shit Done is interesting not only as a repo, but as evidence that agent productivity tools are maturing into process frameworks and inheriting governance questions along the way.

Primary source: GitHub repository. Community discussion: Hacker News.