mtperessoni

AI Starter Kit — AI skill for Claude Code

AI community

Make any repository AI-driven: PRD and TRD as the source of truth, a gate for every change, efficient subagents, related-tests-only workflow and an AI-readable structure, adapted to any stack by Claud.

How to install AI Starter Kit

This entry records only its repository, not the path inside it, so there is no exact command to give. Open mtperessoni/ai-starter-kit and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What AI Starter Kit does

Make any repository AI-driven: PRD and TRD as the source of truth, a gate for every change, efficient subagents, related-tests-only workflow and an AI-readable structure, adapted to any stack by Claude Code skills.

Alternatives in AI

README

AI Starter Kit

[![CI](https://github.com/mtperessoni/ai-starter-kit/actions/workflows/ci.yml/badge.svg)](https://github.com/mtperessoni/ai-starter-kit/actions/workflows/ci.yml)

**Make any repository AI-driven in one command.** The kit installs a PRD and TRD that act as the single source of truth for product behavior, a gate that sends every change through them, subagents that work in parallel cheaply and talk through files, tests that run only for what changed, and a code structure an agent can navigate without reading whole files. It adapts itself to the project's stack.

Every rule here was built, measured and corrected in a production AI service, then extracted so new projects start where that one ended.

Why

AI agents in a real codebase fail in predictable ways: they read whole files to find one rule, they rediscover the domain on every task, they implement from the code instead of from a decision, they run the whole test suite after every edit and then spend hours fixing tests unrelated to their change, and review loops never end.

What the rules in this kit changed, measured in the source repository:

Change Result
Orchestrator conducts, workers write files, the orchestrator reads only their short returns End-to-end time -48%, cost -53%, main context peak from 196k to 76k tokens, same quality
PRD split into one small file per section, with an index, instead of two files of 64k and 165k characters An agent loads the one rule it needs
Related tests only while working, the full suite once at the end against a baseline Tasks that took 13 to 14 minutes because each ran the full suite stopped doing it
Code restructured by feature, size limits enforced by a ratchet Modules over 500 lines from 18 to 1; a 4,056-line class became 23 collaborators
Review capped at 5 rounds, re-review scoped to the previous findings The loop that had consumed 3.6 of 8.8 agent hours in one delivery stopped

How it works

f