CrawlForge presents 31 tools with inconsistent definition quality. While descriptions are present for all tools, they are frequently under-described (most are 20-80 chars), lack actionable guidance for LLM selection, and many tools conflate multiple concerns into single operations. Input schemas are minimal, most tools accept only 1-3 parameters with basic type declarations but no constraints (enums, ranges, patterns). Output schemas are entirely undocumented: the source code shows no response shape definitions anywhere. No tool provides error guidance, parameter relationships, or recovery hints. The 'scrape' tool (tool 6) and 'agent' tool (tool 31) exemplify scope creep, 'scrape' accepts a 'formats' array and returns 'markdown, html, rawHtml, text, links, metadata, branding, screenshot, json, highlights, question', a kitchen-sink design that forces the LLM to reason about format dependencies rather than the server providing structured variants. The 'browser_session' (tool 23) tool accepts an 'action' parameter as free-form string ('create, close, execute') with no enum constraint, inviting hallucination. Naming is generally action-verb compliant (fetch_, extract_, scrape_, search_, list_, etc.), but several tools lack clear differentiation: 'extract_text' vs 'extract_content' vs 'extract_structured' are confusingly similar and lack descriptions explaining when to use each. The codebase shows tool definitions in scripts/tool-sweep.mjs and server implementations in src/, but actual schema registration is not visible in the provided source, tool-sweep.mjs is referenced but not included, making it impossible to verify whether schemas actually exist in the code or are inferred from the summary. This caps per-tool scores at 50 for tools where registration cannot be directly verified.

• Document output schema for all 31 tools. For each tool, define the returned fields, types, and a summary example. Example for extract_links: 'Returns {links: [{url: string, text?: string, title?: string}], total: number}'. Use this template for all tools.
• Expand all descriptions to 100-200 characters and include WHEN-clause guidance. Template: '[ACTION VERB] [RESOURCE] [OUTPUT FORMAT]. Use this when [SCENARIO]. Returns [FIELD1: type1, FIELD2: type2, ...]. Prerequisites: [if any].' Example: 'Extract all hyperlinks from a webpage as a structured list. Use when you need to find related resources or validate link integrity. Returns {links: [...], total: number}. Works with public URLs only.'
• Split or rename similar extraction tools. Consolidate extract_text, extract_content, and analyze_content into a single 'extract_content' tool with a mode parameter (text|summary|analysis). Do the same for extract_structured and extract_with_llm, use a single tool with an 'llm_mode' boolean or 'extraction_method' enum. Reduce cognitive load for LLM tool selection.
• Add enum constraints to all freeform string parameters. browser_session 'action' → enum: ['create', 'close', 'execute']. scrape 'formats' → enum: ['markdown', 'html', 'text', 'links', 'metadata', 'json']. scrape_structured 'selectors' documentation should clarify CSS selector syntax and examples.
• Document and enforce result limits. Add 'Limit: N results per call' to descriptions (recommend 20-50 for lists). Add offset/limit or cursor parameters to extract_links, search_web, reddit_search, crawl_deep, map_site. Return next_cursor or page tokens for pagination.

Inferred effective spec: 2025-06-18+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry