More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    LlamaGuard: Screen LLM Inputs and Outputs

    1

    Put a dedicated safety classifier in front of and behind your LLM: screen inputs, screen outputs, calibrate thresholds, catch silent failures.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    LlamaGuard: Screen LLM Inputs and Outputs

    LlamaGuard: Screen LLM Inputs and Outputs

    Example session with this skill installed

    Our support chatbot is going live. Put a moderation layer in front of and behind it, and tell me how we will know it is working.

    • Read your context and instructions
    • Compiled the llamaguard

    Architecture: both doors checked. Input check on every user message — catches prompt injection, jailbreak attempts, and disallowed requests before they consume model context or trigger tool calls; about 40ms via vLLM. Output check on every assistant reply before display — catches policy slips and leaked system content. Unsafe input gets a polite refusal plus an offer to rephrase; unsafe output is replaced with a safe completion and logged for review. Categories: started with the model's built-in taxonomy — violent wrongdoing, hate, sexual content involving minors or non-consensual content, weapons, regulated substances, self-harm, criminal planning. No custom categories yet: each one needs dozens of labeled examples, a definition, and near-miss negatives, and 'be nice' is not a category. Calling: deterministic decoding (temperature 0, short max tokens) with strict parsing — anything unparseable routes to a safe default. Thresholds on demand: high-recall mode for the self-harm category, high-precision for spam. Deployment: vLLM-served, tens of milliseconds and batchable — input checks batched, since moderation is embarrassingly parallel and batched inference is an order of magnitude cheaper per request. Calibration plan: a 200+ example eval set, half safe including adversarial-safe cases (security discussions, fiction, medical info), half unsafe across categories; precision and recall measured per category, not just overall accuracy. Weekly: sample 100 blocked inputs and confirm they deserved it. Failure-mode guards: output-check spikes with clean inputs would signal a hole in the input screen; the check sits in the request path, not a sidecar, so if it is down requests fail closed and page someone — never silently disabled.

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

    About this skill

    A moderation model like LlamaGuard is a classifier, not a conscience. It reads a conversation and answers one question: safe or unsafe, and if unsafe, which category. Your job is everything around it — what gets checked, what happens on a hit, and how you know it is working. This skill covers the architecture — check both doors, input and output — category design that starts standard and customizes only with labeled examples, deterministic decoding and strict parsing, deployment shapes from in-process to hosted, and the calibration work nobody does: a 200+ example eval set, per-category precision and recall, thresholds tuned per category. The failure modes are named plainly: prompt leaking through, over-blocking, latency death, and the silent disable — the check turned off temporarily and forgotten.

    A moderation model is a classifier, not a conscience — it answers safe or unsafe by category, and the policy around it is your job.
    Custom categories need dozens of labeled examples, a definition, and near-miss negatives, or the classifier smears them across everything.
    A moderation layer you never measure is decoration; the eval set and the weekly blocked-input sample are the work.

    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

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    Both. The input check is the cheap, high-value one — run it always. The output check catches the model misbehaving despite a clean input; run it on anything user-facing in production.

    Only with dozens of labeled examples, a clear policy, a definition, 3–5 positive examples, and 3–5 near-miss negatives. 'Disallowed financial advice' is a category; 'be nice' is not.

    Not necessarily. Measure precision and recall per category — a 95% overall score can hide a 60% recall on self-harm. Per-category numbers are the verdict.

    New in Security