- Home
- Skills
- APIs & Backend
- Content Moderation a Real Team Can Run
Content Moderation a Real Team Can Run
Size moderation thresholds against real volume and reviewer capacity, with appeals, legal escalation and DSA statements of reasons.
$49
Works with the AI Agents you already use
Content Moderation a Real Team Can Run
Example session with this skill installed
Audit my moderation policy for 50,000 daily posts where I want to ban users at 0.95 precision for harassment and flag items for review between 0.7 and 0.95.
- Read your context and instructions
- Compiled the content moderation a
I have analyzed the moderation policy parameters. The current threshold for harassment will lead to significant wrongful actions and queue overflow. ## Policy Triage Audit | Metric | Harassment Category | Analysis | | :--- | :--- | :--- | | Daily Volume | 50,000 items | Base throughput for calculation. | | Auto-Action Threshold | 0.95 | **CRITICAL**: Expected ~10 wrongful bans/day. | | Review Queue Range | 0.70 - 0.95 | **WARNING**: Estimated 1,200 items/day in queue. | | Reviewer Capacity | 8 hours/day | **FAIL**: Queue requires ~25 hours/day at 50 items/hr. | | Appeal Route | Required | Must be routed to a human not involved in initial triage. | | DSA Compliance | Statement of Reasons | Required for all actions above 0.70. | **Policy Recommendations:** 1. Increase `seuil_auto` for permanent bans to 0.995 or change the action to `hide_pending_review`. 2. Narrow the human queue band (e.g., 0.85 - 0.995) to match the 8-hour staff capacity. 3. Implement hash matching for repeat offenders to bypass classification entirely. **Next steps** - Update `policy.json` with a 0.995 threshold for irreversible actions. - Integrate the `triage.py` script into your CI/CD pipeline. - Verify the statement of reasons template against your local legal requirements.
Connects securely to your tools. The creator never sees your data.
About this skill
Thresholds sized against your volume, your team and your appeals. Build content moderation a site can actually run - classifier thresholds sized against real volume and reviewer capacity, a human queue with an appeal route, hash matching and legal escalation for illegal material, notice-and-action and statements of reasons under the DSA, audit logs and transparency reporting. Use when a site opens user-generated content, drowns in reports, bans people wrongly, or must answer a regulator. Two failures, and both are arithmetic before they are policy. A threshold set by intuition bans innocent accounts every day. A queue sized by hope grows until nobody reads it, and the site is unmoderated while believing it is not. Integration is spelled out: the single decision record that serves the user notice, the appeal, the audit and the transparency report at once; the notice-and-action webhook; the classifier call with a timeout that fails open for low severity and closed for the legal categories; the two queue metrics that actually matter (age of the oldest item, items per reviewer-hour); hash matching before classification; and a CI gate on the policy file. Includes an EXPERT.md reference (exact parameters, thresholds, failure diagnosis): the threshold table by action type and reversibility, the DSA duties mapped to fields in the decision record, the route for illegal material, reviewer-welfare controls treated as reliability properties, the weekly protocol for measuring what you miss, audit-log retention, and the failure table. Tested scripts: scripts/triage.py - at 50 000 items a day it shows 200 auto-actions and 0.8 wrongful ones for a reversible removal at precision 0.996, and 7.5 review-hours a day against a capacity of 36. Change the action to a ban at precision 0.95 and the same volume produces about 10 wrongly banned users a day. It refuses a policy that automates an irreversible action at harmful precision, that has no appeal route, that leaves a legally reportable category with no external escalation, or that queues more than the team can read. Acceptable use: moderating content on a service you operate, against published rules, with notice to users and a route to contest. The skill refuses covert moderation that hides removals from those affected, suppression of lawful speech presented as safety, profiling by political opinion, religion, origin, health or sexual orientation, using the pipeline to surveil individuals rather than assess content, and designing away the appeal. Not legal advice: DSA and equivalent duties depend on your size, role and jurisdiction. Have the result reviewed before relying on it.
Requires external classifier or LLM scores. Does not include pre-populated hash lists. User must define their own jurisdictional legal categories.
How to use
- 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
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 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 27 days ago
Requires