CARL ships three reference prompts as the blessed AI-assisted evaluation methodology for the standard. They are versioned alongside the standard (prompt v1.0 is conformant with standard v1.0) and are proprietary to Wentzel.ai, all rights reserved (see License).
The prompts are designed to be used by any LLM client — Claude Code, the hosted assessor app, a custom pipeline — and produce structured output that validates against the published JSON Schemas.
CARL ScanComprehensive codebase audit. Nine phases, every CARL pillar, produces a level and per-criterion results.CARL Review — TeamIndividual contributor evaluation in a multi-developer repository. Twelve dimensions, git-attributed, evidence-required.CARL Review — SoloIndividual contributor evaluation for a solo-authored repository. Same twelve dimensions; git attribution skipped.
How to use them
- Pick the prompt that matches what you're evaluating:
- Scanning a codebase for agent readiness →
Scan. - Evaluating a developer's contributions in a repo with multiple contributors
→
Review — Team. - Evaluating a developer from a solo-authored repository →
Review — Solo.
- Scanning a codebase for agent readiness →
- Paste the prompt into your agent (Claude Code, the CARL hosted app, or your own pipeline).
- Run it against the target repository. Provide any requested context
(developer identifier, evaluation period, scope). For
Review — TeamandReview — Solo, the prompts will stop and ask you to fill in missing fields. - Validate the output against the matching JSON Schema. The CARL
prompt-runner(M7) does this automatically and retries up to three times on malformed output.
Guarantees and limits
Each prompt is:
- Versioned.
v1.0will not change semantics once ratified. Editorial fixes ship as patch bumps. Semantic changes require a new version line (v1.1, v2.0). - Evidence-required. Every scored dimension or criterion must be supported by file paths, line numbers, commit SHAs, or command output. The prompts explicitly instruct the LLM to write "Insufficient evidence" rather than guess.
- Sanitized on output. The JSON schemas reject structured output that contains raw code blocks in evidence fields; a conformant runner rejects malformed output before it reaches the report.
- Not magic. An AI-assisted assessment carries variance. See assessment methodology for the standard's requirements on disclosing variance and the three techniques for reducing it.
Related
- Assessment methodology — assessor types, determinism requirement, applicability marking.
- Report format — the output contract every prompt adheres to.
- JSON Schemas — the strict contract for structured output.