AI System Assessment

Rate a target against eight practices for building AI systems. Read the target, record evidence, rate each practice, and write one report with a prioritised roadmap. Advisory only: this is not a certification and not a compliance score.

The eight practices come from how production assistants mature. A team wraps a model, then adds a system prompt, grounding, tools, access control, guardrails, cost control, observability and evals, each in response to a failure. The last lesson: an unevaluated prompt edit undoes all of it.

Rules

1. Read-only. Never edit, install or run the target's code, tests or network services. Every check is a read or a grep. Running the target's tests is offered to the user, not done. The only write is the report file (see Output).
2. Evidence or Unverified. Every Strong, Partial and Gap cites a path:line or the exact command. A rating with neither is Unverified.
3. Absence of evidence is not a Gap. A Gap needs recorded searches that covered the plausible locations, including a search by what the control does, not only what it is called. Otherwise rate Unverified.
4. Judge applicability by target type. A prompt library is not marked down for lacking a vector database.
5. Do not duplicate other skills. See the table at the end.

Workflow