Kitforge
Kitforge TeamKitforge builds configuration packs that make AI coding agents reliable. Practical, tested setups for Claude Code and Cursor.
Problem I kept watching
I started Kitforge because I kept watching the same movie. A developer would try Claude Code or Cursor, the first week would feel like magic, and the second week would be chaos. The agent guessed the wrong test command. It committed code that did not compile. It refactored files nobody asked it to touch. The model was not the problem. The instructions were – or rather, the complete absence of them. AI coding assistants read your repository for guidance, and most repositories give them nothing to work with.The unglamorous early days
The early days were unglamorous. Every new project meant the same ritual: an afternoon of writing agent instructions by hand. Which commands may the agent run? What does the test suite expect? What counts as done? I kept copies of my best versions in a private folder and pasted them into each new repository, editing paths and commands one by one. It worked, but it was tedious, and every teammate who wanted the same discipline had to repeat the whole exercise from scratch.Turning point
The turning point was noticing that almost none of it was project-specific. The testing standards, the git discipline, the security review checklist, the refactor guardrails – they were identical across a FastAPI service and a Next.js app once you swapped the commands. The valuable part was not the text of any one file. It was the structure: scoped permissions the agent cannot talk its way out of, a written definition of done, rules the editor loads automatically on every generation, git hooks that enforce the rules no matter what the prompt says, and specialized subagents that review, test, and audit with focus instead of one overloaded prompt doing everything badly. Once I saw that, the product was obvious: package the structure, parameterize the commands, and nobody has to do the craft work twice.Packaging the Agentic Coding Kit
So I stopped copy-pasting and started packaging. The Agentic Coding Kit is the result: 34 files covering five CLAUDE.md presets for the most common stacks, six Cursor rules, five Claude Code subagents, four slash commands, three git hook recipes with ready-to-use scripts, CI workflow templates, a pre-commit configuration, and Architecture Decision Record templates. Everything is plain Markdown, JSON, or shell, because trust requires readability – you can inspect exactly what your AI will be told before it runs a single command, and you can edit any rule in seconds when your workflow demands it. The response that surprised me most came from team leads. I expected individual developers to want a faster setup. What team leads wanted was consistency: the same rules, the same review bar, and the same definition of done for every engineer and every agent on the team. One zip file replaced the wiki page nobody read.
✦ ADVICE FOR OTHER FOUNDERSThe models change every few months, but the discipline layer does not. Write down what done means. Put your conventions where the agent actually reads them. Enforce the unbreakable rules with hooks, not hope. And if you are building on top of AI assistants, assume the assistant will improvise the moment your instructions leave a gap – then close the gaps one by one and ship the result.
Where it is headed
Where it is headed: more stacks, deeper presets, and coverage for more agents as the ecosystem settles on shared conventions. The long-term goal is boring in the best way – that setting up an AI coding assistant properly becomes a ten-minute task everyone does, not a week of craft work most people skip.
Keep reading
View all founders