Second brain, das alle Portfolio-Projekte von Marco Stang kennt und Fragen dazu beantwortet — als Chat und als MCP-Server
Two tools with minimal documentation. Both lack parameter descriptions entirely, the rubric requires descriptions for every parameter. Tool descriptions are present but generic (German language, no actionable context for LLM selection). No output schemas documented. No error handling guidance. No input validation hints. Naming is acceptable (verb_noun style: list_*, ask_*) but descriptions fail to explain WHEN to use each tool or what structure is returned.
Beantwortet eine Frage zu Marcos Portfolio-Projekten.
Listet alle bekannten Portfolio-Projekte mit id, Titel und Repo-Pfad.
No input/output schemas documented. list_projects returns list[dict] with no field documentation; ask_about_projects returns str with no structure hint. LLMs cannot plan downstream calls or extract typed data.
Parameter 'question' in ask_about_projects has a description in German ('Die Frage zu den Portfolio-Projekten') but lacks format/length constraints, examples of good vs bad questions, or dependency hints. Descriptions must be actionable for LLM reasoning.
Tool descriptions do not explain WHEN to use each tool or what distinguishes them. 'Listet alle bekannten Portfolio-Projekte' (lists all projects) vs 'Beantwortet eine Frage' (answers a question), but when should an LLM choose one over the other? No discovery guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
No error handling or recovery guidance. If ask_about_projects fails (LLM provider error, snapshot load failure, malformed question), the tool returns nothing actionable. No categorization of errors as retryable, user-fixable, or fatal.
list_projects returns raw dict objects with no pagination, limit, or total count. If projects grow to hundreds, the response will bloat the context window. No guidance on result limits or pagination strategy.