Inspiration
Every automation project I start re-builds the same bones: a scheduler, config loading, state that survives a crash, and a way to notify a human. Frameworks solve this with thousands of lines and a dependency tree. For Kiroween's Skeleton Crew category I wanted the opposite: the smallest honest skeleton a solo builder can read in one sitting — and proof that it is a real template, not a toy, by shipping two genuinely different applications on top of it.
What it does
CrewKit is a lean automation template in pure stdlib Python (no pip installs). core/crewkit/ (~433 LOC): config.py (JSON + env overrides), clock.py (UTC time + a pure due-job scheduler), store.py (atomic JSON state via temp-file + rename), notify.py (console/file/webhook sinks), runner.py (runs due jobs, records outcomes, alerts on failure — a raising job never kills the loop). App 1, apps/telegram-task-bot/: a Telegram task-reminder bot (/add, /list, /done) whose scheduler job nudges you when tasks fall due. App 2, apps/price-watch/: watches product pages, extracts prices from JSON-LD (regex fallback), and alerts the moment a price hits your target. Both apps run in an offline --demo mode — no accounts, tokens or network — and each app folder is self-contained: copy core/ + one app folder anywhere and it runs.
How I built it
Spec-first with Kiro CLI. The repo's .kiro/ directory is the actual build configuration: steering files (product/tech/structure), a written spec (requirements R1–R6, design, tasks), a custom crewkit-builder agent created with kiro-cli agent create, and a hook that re-runs the deterministic self-test on every Python edit. Two Kiro sessions built the core and then both apps, each under a hard rule: finish only when python3 tests/selftest.py passes. Full story in KIRO.md in the repo.
Challenges I ran into
Keeping the core honest: the scheduler had to stay a pure function of (now, state) so demos and tests could use a fake clock; state writes had to be atomic without any library help; and the two apps had to stay strictly on the core's public API — no reaching into internals — or the "template" claim would be hollow.
Accomplishments that I'm proud of
34/34 deterministic tests pass offline in ~0.25 s, including subprocess smoke tests that run both apps' demo modes end to end. The core landed at 433 lines against a 700-line steering budget. Zero dependencies: clone, run, ship.
What I learned
A small spec plus a strict steering budget beats a big prompt. Letting Kiro generate against written requirements — and forcing every session to end at a green self-test — produced code that needed almost no hand-fixing.
What's next for CrewKit
More sinks (ntfy, Slack), a third proof app (RSS digest), and a crewkit new scaffolding command generated from the same spec workflow.