SpecKit alone, I gave up — with an orchestrator, I shipped 64 specs in a row
Turning SpecKit's manual pipeline into a safe autonomous loop — backlog as source of truth, one feature per turn, hard-coded invariants enforced in CI.
SpecKit on its own, I gave up. But with an orchestrator, I shipped 64 specs back to back.
Here's what changes everything.
The problem with vanilla SpecKit
It's a manual pipeline: specify, plan, tasks, analyze, implement. At every step, you're the driver. Great for framing one feature, exhausting across an entire backlog.
The autonomous loop
So I wrote a SKILL.md that turns this pipeline into an autonomous loop. And the game-changer is not "I let the agent go YOLO."
It's the opposite.
Three rules make the autonomy safe:
- The backlog decides. Single, ordered source of truth. The agent takes the first "pending" item, drives it end to end, then STOPS. Never again "which feature?"
- One feature per turn. It does one, stops dead, /loop relaunches it for the next. Fresh context every time. Zero drift.
- Hard-coded invariants. Never merge red: it only merges if lint, typecheck, test and build are green. Otherwise, WIP commit and blocked. A critical analyze? It stops and waits for a human decision.
The result
With this orchestrator, I went from pilot to supervisor. The agent drains the queue, even overnight, stops cleanly at the slightest doubt, and never merges red.
Autonomy isn't letting the agent loose. It's giving it informed defaults, hard bounds, and a clean-context loop.
And you — do your agents wait for orders, or drain a queue?