Gameplan

A gameplan is a structured JSON document, comprising concrete plans for the execution of one milestone.

It combines:

  • Human-readable descriptions of the work to be done, the technical decisions made, and the open questions that remain;
  • Natural-language prompts for coding agents to implement well-defined separate patches;
  • Structured, machine-readable data describing a dependency graph among patches;
  • Formal specifications describing invariants of the resulting work.

Theory is the artifact

Writing a gameplan is a process of collaboration with an LLM; not a prompt to be one-shotted.

Using the /write-gameplan prompt, the engineer works with the model to hammer out a concrete shared understanding of the work to be done and how to do it.

This is the opportunity to eliminate open questions and ambiguity. The engineer walks away with a solid understanding of the solution, at an appropriate (architectural) level of abstraction. And the gameplan contains enough detail for a fleet of agents to execute it independently.

Structure for parallelism

Armed with a concrete understanding of the work to be done, the LLM can break it up into a set of well-defined patches, each with its own changes, each making a single well-formed commit.

The LLM will analyze the patches for dependency, and include a dependency graph within the gameplan. This will describe the set of patches that can be executed independently at any point in time.

Structure for mergeability

Let us define mergeability as the ability of a patch to be merged without human review.
Organize patches to front-load and separate out the most mergeable code.
Ensure that patches requiring human review contain minimal noise and maximal context.

Structure for correctness

In addition to the workstream specification, the LLM will construct type-checked, model-checked Pantagruel specifications for the gameplan and for each patch.

Coding agents and reviewers alike can use these as unambiguous specifications of implementation correctness.

Read more