Single tool 'adr-index' has a detailed description (194 chars, within baseline 34-392 range) explaining dual modes and use cases. However, the input schema exhibits critical structural issues: parameters are split across three fields (adr_dir, adr_dir_opt, adr_dir_arg) creating ambiguity and mutual exclusivity that is not documented in parameter descriptions. Most parameters lack type information in the visible schema (adr_dir, adr_dir_opt, readme are strings but lack format/pattern constraints). The schema does not document output structure, pagination, or error cases. Parameter descriptions are present but do not explain dependencies (e.g., 'adr_dir_arg selects README mode unless -o/--adr-dir is used' is buried in tool description, not in param descriptions). No error handling guidance visible. Tool name 'adr-index' is a noun phrase, not a verb_noun pattern (should be 'generate_adr_index' or 'build_adr_index'). The tool performs multiple distinct operations (context mode vs README mode) which violates single-responsibility principle.

• Rename tool to 'generate_adr_index' (verb_noun pattern) to signal action intent to LLM.
• Add explicit type, format, and constraint definitions to all parameters in schema. Use enum for 'format' in schema, not just description. Add minLength/maxLength for path parameters.
• Document parameter dependencies in each parameter's description: 'adr_dir_arg: Positional argument for README mode. Mutually exclusive with --adr-dir and --output flags.'
• Add output schema documenting return type for each format (markdown string, JSON object with fields, graph JSON structure). Include example output.
• Split into two tools: 'generate_adr_context_index' (context mode) and 'update_adr_readme_index' (README mode) to follow single-responsibility principle and reduce cognitive load on LLM.