Reverse engineer clean behavioral specs from any codebase.

Greenfield reads source code, documentation, SDKs, runtime behavior, and binaries, then produces behavioral specifications, test vectors, acceptance criteria, and a full provenance trail. The output describes what the software does — not how any particular codebase does it — so a fresh implementation team can build against the specs without inheriting the original's internal structure.

Greenfield is a Claude Code plugin. It runs inside the claude CLI and uses Claude agents as the workers that read code, write specs, and audit output.

/plugin marketplace add prime-radiant-inc/prime-radiant-marketplace
/plugin install greenfield@prime-radiant-marketplace

Restart Claude Code after installing.

claude
> /analyze /path/to/target

/analyze runs a seven-layer pipeline. It discovers the available intelligence sources (source, docs, SDK, community, runtime, binary, git history, tests, UI, contracts), gathers evidence from each, synthesizes behavioral specs with provenance citations, generates test vectors and acceptance criteria, sanitizes the specs of implementation details, then runs a second-pass review of the result.

Target shape doesn't matter. Greenfield has been used on single-file minified JavaScript bundles and on source-tree projects; the pipeline also has a path for decompiled native binaries. The methodology adapts to what's there.

Workspace output:

workspace/
├── raw/         # Analysis artifacts with source references
├── output/      # Sanitized specs for the implementation team
│   ├── specs/
│   ├── test-vectors/
│   └── validation/
└── provenance/  # Citation audit trail

claude
> /sanitize /path/to/workspace

Re-runs the sanitization pass on an existing workspace. Useful when the initial pass left contamination that the audit caught.

Greenfield stops at the specs. The implementation team reads workspace/output/ and builds against the behavioral specs, test vectors, and acceptance criteria there.

The pipeline dispatches agents with role-specific prompts across the seven layers. Each dispatch runs one of two generic agent types loaded with a specific skill:

• analyzer — the workhorse. Reads evidence and writes specs. Dispatched under many roles across layers: discovery-agent, bundle-splitter, chunk-analyzer, function-analyzer, doc-researcher, community-analyst, sdk-analyzer, cli-explorer, behavior-observer, feature-discoverer, architecture-analyst, api-extractor, synthesizer, module-mapper, deep-dive-analyzer, behavior-documenter, user-journey-analyzer, contract-extractor, spec-verifier, source-completeness-checker, test-vector-generator, test-generator, acceptance-criteria-writer, spec-reviewer, structural-leakage-reviewer, content-contamination-reviewer, behavioral-completeness-reviewer, deep-read-auditor, fidelity-validator.
• sanitizer — rewrites raw specs into output specs free of implementation details. Dispatched during Layer 5 and for remediation during Layers 6 and 7.

Apache 2.0. Copyright 2026 Prime Radiant, Inc.