Loading ToolBench
MCP server exposing the 90-day AI transformation method as tools: score departments with evidence-backed scores, inspect the portfolio, draft decision memos, and run governance red-line checks.
Transformation OS demonstrates solid tool design with clear naming (verb_noun pattern), comprehensive descriptions (avg 180 chars), and well-structured schemas. All 10 tools have explicit input schemas with typed parameters and descriptions. Evidence-driven scoring (submit_department_scores) enforces data integrity. However, output schemas lack formal documentation, error handling is minimal, and some parameter descriptions could be more prescriptive about constraints and formats. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Deterministic governance red-line check for an AI use-case idea. data_classes: any of public, internal, confidential, personal, employee. Returns blocked / review_required / fast_track with reasons and required approvals.
Four-eyes calibration: compare two independent scorings of the same department. Flags dimensions diverging >= 2 points (must be discussed with evidence), shows whether the disagreement flips the portfolio quadrant.
Markdown skeleton for the day-30/60/90 executive decision memo (result / plan / budget+resolutions), prefilled with the live portfolio.
The 45-minute lead-interview guide (German): 12 questions in 4 blocks, with a debrief checklist. Use it to conduct or debrief an interview.
The 6-dimension scoring rubric (anchors per dimension) and the scoring rules. Read this before scoring a department.
The assistant view for a program day: current phase, this block's to-dos (with template paths), next milestone, phase warning signals, and the live program state (departments scored, recommended pilots) so gaps are visible. Omit program_day to compute it from the start date set via set_program_start.
Output schemas not formally documented. Tools return dicts/strings but LLMs cannot see field structure, forcing them to infer response shape and risking extraction errors.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools like submit_department_scores and sync_outputs modify state but lack explicit risk markers for agent planning.
Error handling minimal. submit_department_scores rejects fabricated quotes but returns no recovery guidance. sync_outputs subprocess call has no timeout error handling or actionable failure messages.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Current portfolio: every department with impact, readiness, quadrant and role, plus recommended pilots and the deliberately-postponed list.
Set the program's day-1 date (ISO format, e.g. 2026-09-01). Stored in org.yaml so get_todos can compute the current program day automatically.
Submit scores for a department (all 6 dimensions, integers 1-5). EVERY score requires evidence: a verbatim quote from the interview, a metric, or a system inspection (>= 20 chars per dimension). Submissions without evidence are rejected — that is the method's core rule, enforced. ALWAYS pass the interview transcript/notes as source_text when you have them: every quotation-marked span in the evidence is then verified VERBATIM against the source (whitespace/case-insensitive) — fabricated quotes are rejected. Evidence not from the source (a metric, a system inspection) must be prefixed 'metric:' or 'system:'. Persists to org.yaml; call sync_outputs afterwards.
Regenerate dashboard data and score tables from org.yaml (runs tools/generate.py).
Parameter constraints under-specified. check_use_case data_classes accepts 'list | None' but lacks enum constraint; day parameter in draft_decision_memo accepts any int without validation (should be 30|60|90).
Pagination not implemented. list_portfolio and get_todos return unbounded results; no limit, offset, or cursor parameters. Large portfolios could exhaust context window.