Turning coding agents from improvisers to engineers

What it does

This repo is a collection of six agent “skills” — folders containing SKILL.md instruction sets that Claude Code, Cursor, and Codex invoke when a task matches their description. Together with an AGENTS.md template, they enforce a four-phase workflow: isolate new work in Git worktrees, build to a service-layer architecture, prove changes with video or screenshot evidence, and ship only after automated review loops return a clean bill of health. It is essentially a playbook designed to replace ad-hoc agent improvisation with repeatable process.

The interesting bit

The standout is evidence-driven-testing, which bundles a cross-platform Python recorder that drives the UI via computer-use tools, burns timestamped annotations into an evidence.mp4, and generates a report.md — turning “trust me, it works” into verifiable artifacts. The new-feature skill uses Git worktrees so multiple agents can work the same repository in parallel without collision.

Key highlights

• new-feature isolates every task in a fresh worktree branched from origin/main, complete with scope checks and post-merge cleanup, built for concurrent agent sessions.
• code-structure enforces a two-layer architecture — actions handle domain rules (“why/when”), services handle reusable mechanics (“how”) — and includes a migration checklist plus anti-pattern tables.
• evidence-driven-testing captures MPEG-TS streams on Linux, macOS, and Windows; even a crashed recorder yields usable footage, while headless environments fall back to Playwright screenshots.
• greploop iteratively fixes PRs against Greptile reviews until confidence hits 5/5 with zero unresolved comments, up to ten cycles; greploop-apps provides a fallback for oversized PRs that exceed file-count limits.
• before-and-after generates PR-ready visual diff tables, vendored from Vercel Labs.

Caveats

• Several skills are vendored from upstream projects (before-and-after from Vercel, greploop from Greptile), so the repo functions partly as a curated distribution with local variants.
• The evidence recorder demands FFmpeg built with libx264 and ass filter support, and on Linux it only supports X11 or wlroots Wayland — GNOME and KDE are explicitly excluded.
• Most of the project lives in markdown instructions rather than executable code; only evidence-driven-testing ships substantial automation scripts.

Verdict

Best for teams running multiple AI agents on production codebases who need architectural consistency and audit trails. Skip it if you are looking for a standalone CLI tool — this is organizational plumbing, not an app.