0022. Functionalize Repeatable Work Into Deterministic Surfaces

Status

Accepted

Context

NEXUS is being developed with help from multiple AI systems and humans.

That creates a real risk:

  • useful work patterns can remain trapped inside the habits or capabilities of one particular AI
  • repeated tasks can stay ad hoc instead of becoming inspectable system behavior
  • later collaborators may get different results because they are relying on model improvisation instead of stable reviewed surfaces

NEXUS is explicitly trying to move in the opposite direction.

When a task pattern becomes important and repeatable, the goal is to turn it into something more concrete:

  • a function
  • a tool command
  • a script
  • a workflow
  • a schema
  • a deterministic transformation surface

That keeps the work reviewable, teachable, and improvable over time.

This is part of the emerging direction behind FORGE, and it also fits interaction-heavy tool lines such as Penpot-related work.

Decision

Prefer functionalizing important repeated work into deterministic and reviewable surfaces instead of leaving that work dependent on the capabilities of a particular AI model.

Working rule:

  • let AI help discover and prototype the work first
  • once the pattern proves useful and repeatable, capture it in a stable surface
  • prefer surfaces that can be inspected, reviewed, tested, and improved by later collaborators
  • treat ad hoc AI execution as a starting point, not the desired long-term endpoint, when the task is important enough to recur

Examples

Examples of the kinds of surfaces this points toward:

  • CLI commands
  • importer adapters
  • repo bootstrap workflows
  • repeatable extraction or publishing scripts
  • reviewed transformation code
  • deterministic schema-driven conversions
  • compiler-like generation behavior when the rules are stable enough

Examples of areas this should influence:

  • FORGE
  • Penpot interaction and integration work
  • cross-repo bootstrap and maintenance workflows
  • ingestion transforms that begin as exploratory AI work and later become stable pipelines

Consequences

Positive:

  • future humans and AI systems can use the same reviewed surfaces
  • behavior becomes less dependent on a specific model's strengths or quirks
  • repeated work becomes easier to evaluate, compare, and improve
  • deterministic and testable behavior grows over time

Tradeoffs:

  • some exploratory speed is traded for formalization effort
  • collaborators must decide when a pattern is mature enough to encode
  • not every useful one-off should become a framework too early