Don't Write Prompts, Write Software
A presentation at AI Native DevCon 2026, London in June 2026 in London, United Kingdom by Baruch Sadogursky and Macey Baker
Abstract
Skills are more than reusable prompts — they are composable units of behaviour. In this workshop we take a real, end-to-end agent workflow and turn it into a working system built from skills: identifying the right behavioural units, turning them into discrete skills, refining and optimising them with Tessl’s review and eval tools, and bundling them into a reusable plugin. Along the way we explore why this approach is often more robust than relying on a single giant entry prompt, and how modular skills make agent behaviour easier to test, improve, and reuse.
Resources
- Workshop repository — the live build from the session
- Stop Prompt Hacking — Macey’s research: threatening the model, persona cosplay, and tip bribes don’t actually help
- Claude Code — skills, plugins, hooks, and the agentic CLI
- Model Context Protocol (MCP) — the open standard for tool/context integration
- Tessl — author, review, and eval context artifacts
- Tessl registry — publish and install plugins across teams
- jbaruch/coding-policy — the meta-policy plugin the agentic reviewer enforces
- GitHub Agentic Workflows (gh-aw) — agentic reviewers that run in GitHub Actions
- GitHub Copilot code review — the bot reviewer summoned in the demo
- OpenRouter — one API across many models; send cheap tasks to cheap models
- LiteLLM — unified LLM gateway with routing, fallbacks, and cost control
- More talks by Baruch — slides, videos, and shownotes