> *TL;DR:* On larger features, Claude’s plans tend to get vague, which leads to vibe-coded spaghetti in the implementation. This plugin forces PRD + design + task-level acceptance criteria to keep things grounded.

Hi HN,

I use Claude Code a lot and it's native plan mode works well for small to medium changes, but on larger features the plans would be vague, which lead to more hallucinations during implementation.

I ended up coming up with a manual workflow for larger features where I'd write a design first by bouncing ideas back and forth with ChatGPT/Claude/Gemini until I had something broken down into clear phases. I’d drop those docs into the repo and then have Claude Code implement each phase individually while referring back to them.

So I made ShipSpec, a Claude Code Plugin that automates this workflow inside Claude Code. It generates three repo-local markdown files:

* PRD.md – requirements * SDD.md – design doc * TASKS.md – ordered tasks with acceptance criteria

Because tasks link back to both the requirements and the design, Claude Code stays grounded during implementation.

Usage:

``` /shipspec:feature-planning "add user authentication with OAuth2" /shipspec:implement-feature auth-feature /shipspec:implement-task auth-feature TASK-001 ```

Not sure if this is generally useful or just scratches my own itch, but curious if others have hit similar issues with having Claude Code implement larger features.