# Seedkit — cookiecutter as a Claude Code skill, trained in a self-improvement loop

**URL:** <https://forum.djangoproject.com/t/seedkit-cookiecutter-as-a-claude-code-skill-trained-in-a-self-improvement-loop/45113>\
**Category:** Packages\
**Created:** [May 15, 2026, 1:24pm UTC](https://forum.djangoproject.com/t/seedkit-cookiecutter-as-a-claude-code-skill-trained-in-a-self-improvement-loop/45113 "2026-05-15T13:24:51Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![kmmbvnr](https://sea2.discourse-cdn.com/flex026/user_avatar/forum.djangoproject.com/kmmbvnr/32/25527_2.png) [@kmmbvnr](https://forum.djangoproject.com/u/kmmbvnr)\
**Post date:** [May 15, 2026, 1:24pm UTC](https://forum.djangoproject.com/t/seedkit-cookiecutter-as-a-claude-code-skill-trained-in-a-self-improvement-loop/45113/1 "2026-05-15T13:24:51Z")

</div>

I’ve been working with Django since the 0.9x days, and over time modern projects started needing more setup around the framework itself. Cookiecutters and starter templates tried to help with this, and now many people also generate projects with AI agents — which introduces a different set of problems.

I’ve been experimenting with two approaches for self-improving agent skills:

`/seedkit` starts with framework/package docs embedded into the skill, then iterates through run → error → review → fix cycles (plus some manual cleanup to keep it compact).

`/seedkit-slim` starts only with questions about what should be added to a fresh `startproject`, and gradually grows by recording and reusing fixes for its own mistakes.

You can check them at [https://github.com/RobustaRush/seedkit](https://github.com/RobustaRush/seedkit), along with baselines (Claude Sonnet without the skill) and sample generated projects.