Get started / Field Manual
Field Manual
The fastest way to learn Becky. In one sitting you'll install it, run your first build, meet all 16 commands, choose between autopilot and hands-on gating, learn to read the verdict, and see how it ships docs, announcements, and clean commits. No servers, no accounts — a Node toolkit plus a folder of agents.
Your first build
A worked example, start to finish. You type two commands; the crew does the rest. Here's exactly what goes in and what comes out.
first build
$ becky greenfield "my feature"
created phase folders + _task.yaml
$ becky autopilot
[1] Vision research.md ok
[2] Fury brief.md ok
[3] Coulson prd.md + stories ok
[4] Shuri ux-spec.md ok
[5] Strange architecture.md ok
[6] Stark code + tests ok
[7] Loki review-findings ok
[8] Widow test-report.md ok
[9] Heimdall verdict: DONE ok
[10] Watcher wiki updated ok
What you type. becky greenfield "my feature" creates a task — a set of numbered phase folders and a _task.yaml that tracks where things stand. Nothing has run yet; you've just told the crew what you want.
What you get. becky autopilot runs every remaining phase in order. Vision researches the landscape, Fury writes the brief, Coulson turns it into traceable requirements and stories, Shuri and Strange design it, Stark builds it with tests, Loki and Widow try to break it, Heimdall files a verdict, and Watcher records the lesson in the wiki. You come back to finished work with a paper trail at every step.
Prefer to drive? Swap autopilot for run → approve and you get a checkpoint between every phase. Same pipeline, your hand on the wheel. See both modes ↓
Tip: running on an existing codebase instead of a blank slate? Use
becky brownfield "…" — it starts with archaeology, mapping what's really there before anyone touches a line. Full step-by-step in the
Workflow Map.
Autopilot vs manual
Every phase is self-contained — it reads its inputs from a known place and writes its outputs to a known place, with no ambient context needed. That single property is why you get to choose your level of control.
Hands onManual — gated
run executes the current phase, then stops. You inspect the output. approve opens the gate and advances; revise "feedback" sends notes and re-runs the same phase. The gate stays shut until you're satisfied — a checkpoint between every step.
Hands offAutopilot — unattended
autopilot runs all remaining phases in sequence without stopping. Type it before bed; wake up to completed work, every phase documented and every output filed. Because each phase is self-contained, the whole pipeline runs without you.
run → approve → revise
$ becky run
Stark code + tests done
gate: tests pass · clean — awaiting you
$ becky revise "add an empty-state"
Stark re-running build... done
$ becky approve
gate passed → advancing to Loki (review)
Rule of thumb: drive manually the first few runs to build trust in the gates, then hand it the night shift with autopilot. You can always stop and inspect any phase's output — it's all plain text on disk.
Docs & announcements
Once Heimdall signs off, the two wordsmiths step in. Docs are written to match the real code; announcements are anchored to the actual diff. Nothing gets claimed that the code can't back up.
"If a newcomer can't succeed in five minutes, the doc isn't done."
Parker reads the real code and the verdict, then writes the README, the guide, the onboarding, and the comments — clear enough for a stranger to follow. He never documents a feature the code doesn't actually have.
Quill
DevRel & announcements"No vaporware — every note rides a change Heimdall already signed off."
Quill turns the diff into a changelog, a release note, a launch post — the line that makes a developer stop scrolling. Big energy, zero overclaiming; every word is anchored to a real, shipped change.