More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Tool Description Lint — Why Your Agent Picks the Wrong Tool

    1

    Reads your tool definitions and reports the missing boundaries, undocumented parameters, overlapping tools and unguarded destructive actions that make an agent choose wrong.

    Free

    Security scanned
    Tool Description Lint — Why Your Agent Picks the Wrong Tool

    Tool Description Lint — Why Your Agent Picks the Wrong Tool

    Example session with this skill installed

    python3 scripts/tool_lint.py tools.json --fail-on high

    {"name":"send_email","description":"Send an email to a customer.","input_schema":{"properties":{"to":{"type":"string"}}}}

    • Read your context and instructions
    • Compiled the tool description lint

    tool-description-lint 20 tools
    score 5/100 grade F findings 94
    high 5 medium 66 low 23

    HIGH T1 missing description process
    HIGH T5 destructive without approval send_email
    MEDIUM T12 overlapping tool search_web
    MEDIUM T7 undocumented parameter update_status

    FAIL findings at or above high

    Connects securely to your tools. The creator never sees your data.

    About this skill

    When an agent calls the wrong tool, the usual fix is a better prompt. The usual cause is the tool definition. Two tools sound alike, a parameter has no description, a status accepts any string, and a destructive action looks exactly like a read. The model is choosing from what it was given.

    This reads the definitions you send to the model and reports the defects that make selection a guess:

    • Missing or thin descriptions, with a check for an explicit "use when" boundary
    • No exclusion, so the nearest wrong choice is never named
    • Undocumented parameters and unconstrained choice fields that should be enums
    • Generic names such as process, run and get that collide with every neighbour
    • Overlapping tools whose names and descriptions cover the same job
    • Destructive actions with no approval or reversibility stated
    • Duplicate names that silently shadow one another
    • Tool-count pressure: 20+ in one context, and 30+ as a hard flag

    Each finding names the tool, the parameter and the fix, written as a description change rather than "be clearer". Exit 1 at a severity threshold for CI, so a tool registry cannot lose its boundaries in a refactor.

    The shipped failing set is 20 tools with 94 findings, including a duplicate pair of user lookups, a search_web/web_search collision, and a send_email with no approval path. The clean set is six tools with a score of 100 and no findings.

    Good to know: this is static analysis of names, descriptions and schemas. It cannot prove the model will choose correctly; evaluate tool choice on real requests. Overlap findings are review prompts, because semantic overlap depends on the task. Tool-count thresholds follow published guidance, not a measured cutoff for your model. Anthropic reports degradation past roughly 30-50 tools and recommends tool search from about 10; OpenAI suggests fewer than 20 available at the start of a turn.

    Static analysis only. Cannot prove model selection accuracy. Overlap findings require manual review. Tool count thresholds use general provider guidance.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Permissions

    Terminal / Shell
    Network Access
    Read Files
    Write Files
    Listed1 day ago

    What's inside

    Frequently Asked Questions

    This skill identifies 'description defects'—vague or overlapping tool definitions—that cause models to select the wrong tool, hallucinate parameters, or execute destructive actions without proper guards.

    It is compatible with any agent framework using OpenAI or Anthropic tool formats, as well as generic JSON schemas, making it ideal for LangChain, CrewAI, or custom implementations.

    Unlike LLM-based reviews which can be inconsistent, this tool uses deterministic static analysis to provide a reproducible audit trail and CI-compatible exit codes for automated deployments.

    You will receive a Python 3.10+ CLI script that runs locally with zero external dependencies, allowing you to integrate it directly into your existing CI/CD pipelines or local development workflow.

    The linter currently scans for 13 specific defect types, including semantic overlaps, undocumented fields, unconstrained strings, and missing approval paths for high-risk operations like payments or deletions.

    New in LLM Ops