/write-gameplan

This is how Flowglad does vibe-coding.

It's the composition of a couple coding agent skills with a deterministic agent supervisor, onton, all of which participate in a single shared abstraction: the gameplan.

1 w = /write-workstream 2 /write-spec optional 3 for m in w.milestones: g = /write-gameplan m onton --gameplan g
1

write a workstream

Define the vision, current state, key challenges, and milestones. The output is a workstream whose milestones become the unit of iteration.

Read the workstream guide
2

write a spec

Optionally: formally specify invariants that should hold when the feature is complete.

Read about the Pantagruel specification language
3

write and execute gameplans

  1. Turn a milestone into a concrete gameplan.
  2. Execute the gameplan with the onton supervisor.

Repeat for the next milestone in the workstream.

Gameplan guide · Execution guide

Push human intervention to the left

We're all tech leads, now: the gameplan phase is the moment that the engineer sets the technical direction of the feature.

The engineer works with the model to thrash out any open questions and to make explicit all the architectural decisions in the feature.

The gameplan patch dependency model prioritizes parallelism and mergeability.

The less human code review a patch needs, the more mergeable it is. So gameplans front load tests and refactors and ensure that user-observable changes are behind a feature flag.

Use aggressive automated code review, opinionated linting, and exhaustive testing to give the model the maximum mechanical feedback.

The onton supervisor will babysit patches, ensuring that review is addressed, CI failures fixed, and merge conflicts are resolved.

After everything is done mechanically that can be, the engineer merges the patch, automatically kicking off any downstream patches.

Read more