Decomposed ticket workflow from the AI Native DevCon London 2026 context workshop: skills for code and documentation tickets, the mandatory tests-first / clean-review / human-review rules, and the deterministic scripts/ they lean on.
70
88%
Does it follow best practices?
Impact
—
No eval scenarios have been run
Passed
No known issues
Don't Write Prompts, Write Software — Baruch Sadogursky & Macey Baker.
A live, build-it-with-the-room session on treating context as software: take a working but bloated "uber-prompt", decompose it by determinism into skills (conditional behaviour), scripts (the deterministic parts, with real tests), and rules (the mandatory checks), then package the whole thing as a distributable plugin — and apply the same discipline to the context artifacts themselves, with evals as their tests.
The artifacts get built live during the session; this repo is where they land, so you can follow along or take them home.
Slides & resources: https://speaking.jbaru.ch/aidevcon-london-2026-dont-write-prompts
To follow along live (or reproduce the build later) you'll need:
gh CLI —
authenticated, for the issue → PR flow we drive from the terminal.pytest — the deterministic parts get pulled out
into scripts/, which are real Python with a real test suite
(python3 -m pytest scripts/tests).The artifacts are built live, so the layout fills in as the session progresses. Where to look:
skills/ — the conditional behaviour, decomposed from the uber-prompt
into discrete skills.scripts/ — the deterministic parts, pulled out as real code with tests.rules/ — the mandatory checks the workflow must enforce.