Skip to content

AI-Assisted Development

Superloom treats AI agents as a permanent development audience: code and documentation are structured so agents produce conforming work, and the standards for configuring, instructing, and dividing labor with agents are documentation in their own right. This section holds those standards. They are vendor-neutral: the same files and conventions work across agent tooling generations.

Documents

DocumentCovers
Agent ConfigurationThe AGENTS.md standard, tool-specific folders, the size budget, module-level ROBOTS.md
Workflow AuthoringHow to write workflows agents execute with precision: phases, gates, evidence, convergence
Workflow ArchetypesLanguage-agnostic patterns that concrete workflow families are compiled from
Model TieringDividing work between expensive reasoning models and inexpensive execution models

The Operating Ideas

Three ideas run through this section:

  • Ambient rules and procedures are different things. An agent editing existing code needs standing knowledge (AGENTS.md, loaded every conversation). An agent performing a lifecycle operation (create, fix, audit, publish) needs a procedure (a workflow, loaded on invocation). Splitting by access pattern keeps the standing context small and the procedures complete.
  • Execution content is embedded, rationale is referenced. Agents execute reliably from content inside the file they are running, and unreliably across links. Workflows therefore embed everything execution-critical, and a compile step keeps the embedded copies synchronized with their canonical sources.
  • Evidence over trust. Agents claim completion optimistically. Every standard in this section demands visible evidence (quotes, tables, verdict lines, convergence statements) precisely because unverifiable claims are the primary agent failure mode.

Released under the MIT License.