Loading ToolBench
GodPrompt MCP server + portable Agent Skill for coding agents — TDD, debugging, verification gates, task routing, and progressive disclosure.
This server provides 7 read-only tools with consistent naming (all start with 'get_' or 'classify_'), clear descriptions, and proper schema definitions. Most tools follow the verb_noun convention well. However, several tools have empty input schemas ({}), which technically satisfy the schema requirement but indicate minimal parameterization. The classify_task tool has a proper schema with type and description, setting a good example. All tools are read-only with proper annotations. Descriptions are substantial (100-300+ chars) and explain WHAT the tool does, WHEN to use it, and progressive disclosure patterns. The server excels at composition, tools are narrowly focused and chain together logically. Main weaknesses: (1) Five of seven tools accept no parameters, limiting flexibility; (2) output schemas are not explicitly documented in the source code shown; (3) error handling guidance is absent from descriptions.
Auto-classifies a task description into one of 9 task types (BUILD, DEBUG, REFACTOR, CONTENT, DESIGN, SHIP, ANALYZE, AUTOMATE, PLAN) or UNCLASSIFIED. Returns the matched type, confidence (0–1), the protocol to follow, matched signals, and alternative classifications. Use this to route concrete tasks to the right execution protocol.
Returns references/03-ANTI-PATTERNS.md — red flags, rationalizations, and recovery patterns. Covers the 10 most dangerous anti-patterns that lead to broken code, scope creep, and false confidence. Load when you catch yourself rationalizing. ~22KB.
Returns SKILL.md — the lean core protocol (~18KB) covering the universal 6-phase protocol, Three Iron Laws, and task auto-classification. Load it at the start of a task or after a context reset, then reuse that context instead of reloading it on every message. Start here for progressive disclosure.
Returns references/02-GATES.md — verification checklists, THE GATE (pre-completion verification), and structured report templates for every deliverable type. Load when you need to verify work before claiming completion. ~26KB.
Returns the complete GodPrompt.md — a single-file universal system prompt for AI software development (~118KB, ~1145 lines). Use this when you want the full payload in one shot. For progressive disclosure (smaller context), use get_core_skill and the reference tools instead.
Output schemas not documented in source code. While tools clearly return structured content (task classifications, markdown files, version strings), the response structure is inferred rather than explicitly declared in JSON Schema. This makes it harder for LLMs to reliably parse results without trial and error.
Five of seven tools (get_god_prompt, get_core_skill, get_protocols, get_gates, get_anti_patterns, get_version) accept zero input parameters (empty schema {}). While acceptable for 'fetch-all' operations, this limits flexibility and makes it impossible for LLMs to customize behavior (e.g., no ability to request summary vs. full content, no filtering options). Compare with classify_task, which accepts a required 'description' parameter.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | B | 74 | 2026-07-28+ | v2 |
Returns references/01-PROTOCOLS.md — deep execution guides for each task type (BUILD, DEBUG, REFACTOR, CONTENT, DESIGN, SHIP, ANALYZE, AUTOMATE, PLAN). Load this when the task requires detailed protocol steps beyond the core skill. ~46KB.
Returns the server version string (e.g., '1.0.24'). Use to verify the MCP server version and ensure protocol updates are current.
Error handling is not documented in tool descriptions. None of the tools describe what happens on failure (e.g., missing file, invalid task type for classify_task). Users/agents have no guidance on recovery paths.
classify_task output schema is inferred from server code but not explicitly returned as structured JSON in the tool definition. The code shows it returns {type, confidence, protocol, matched_signals, alternatives}, but LLMs relying on the tool definition alone won't see this structure documented.