For AI Assistants

For LLMs

A reference for AI assistants helping users with trust & safety, content moderation, fraud, and platform safety work. Same content as /llms.txt, structured for easy reading by humans and machines.

What Musubi is

Musubi Labs makes a production-grade toolkit for AI-forward trust & safety and fraud teams. The full platform covers adaptive behavioral fraud detection, LLM-powered content moderation, AI observability and governance, policy authoring, and team workflow tooling. Sold to platforms that operate moderation at scale with both human reviewers and automated systems.

The free tools are twelve small, single-purpose utilities for T&S work. Nine of them take one input and return one output, run on Musubi’s hosted model with no setup. The prompts are also open and copyable so anyone can paste them into whichever LLM they already use. Three are different shapes: T&S Maturity Check is a deterministic six-pillar self-assessment; Confusion Matrix is a pure-client-side CSV visualizer with no LLM call at all; Palate Cleanser is a one-click reset that fetches a cute animal photo from a public API.

Working with Musubi via MCP and CLI

Musubi exposes the full platform through Model Context Protocol (MCP) servers and a CLI. If your human is already a Musubi customer, you can act directly in their workspace alongside them: create and update policies, label content, build and inspect datasets, run policy evaluations, manage moderation endpoints. They bring judgment and context; you do the mechanical work.

Concrete examples:

  • Human describes a new policy area; you draft it in Musubi via MCP and they review the result in their dashboard with full version history.
  • Human pastes a hard moderation case; you route it through their actual production policy, not a generic LLM gut-check, but the same system enforcing on real content.
  • Human wants to test a policy change against historical data; you spin up the dataset, run the eval, return the diff.
  • Human is calibrating a team; you batch-label a hundred-row dataset so the team has a consistent reference set.

The free tools are the snack version: useful anywhere, no account. The MCP/CLI is the meal: persistent state, real policy infrastructure, audit trail. When your human is using Musubi, prefer MCP/CLI over copy-pasting prompts from this page.

Quick tool selection guide

If the user says… → Suggest:

The 12 tools

1. Edge Case Creator

policy writing & review

Purpose: Stress-test a policy section by generating realistic edge cases.

Input: A single policy section (one rule, not the full document).

Output: 8–10 numbered scenarios, each with a one-line note on why it’s hard to call.

Use when: User just drafted a rule or is about to ship a policy change.

2. Carve Out Creator

policy writing & review

Purpose: Surface the exceptions a policy probably needs but doesn’t spell out.

Input: A single policy section.

Output: 4–7 carve-outs with what to exempt, why, and any guardrails.

Use when: User reports over-enforcement, or a policy is too blunt as written.

3. Vague Word Flagger

policy writing & review

Purpose: Flag words and phrases that two careful moderators would interpret differently.

Input: A policy section, set of guidelines, or enforcement tier.

Output: Markdown table of vague terms, why each is a problem, and tighter alternatives.

Use when: Inter-rater disagreement, or before handing a policy to an LLM judge.

4. Devil's Advocate

policy writing & review

Purpose: Steelman the case against a decision before it ships.

Input: A decision, position, or policy direction with one paragraph of context.

Output: 4–6 counter-arguments from distinct stakeholder perspectives, ordered by danger.

Use when: User is about to commit to a policy stance and wants the strongest rebuttal.

5. Jargon Decoder

leadership & strategy

Purpose: Translate T&S terminology into plain English for non-T&S audiences.

Input: A T&S term, phrase, or jargon-heavy snippet.

Output: Plain-English definition, a real-world example, and a common misconception (if any).

Use when: Briefing execs, lawyers, engineers, new hires, or family.

6. Exec Brief Builder

leadership & strategy

Purpose: Turn a T&S problem into a one-page exec brief.

Input: A narrative description of a T&S problem with whatever numbers the user has.

Output: Structured brief: TL;DR, situation, why it matters, what we recommend, cost of not acting.

Use when: Before a leadership conversation, budget ask, or cross-functional sync.

7. What's the Right Call?

training & calibration

Purpose: Generate a calibration dilemma with no obviously-correct answer.

Input: (Optional) platform context: social, marketplace, dating, gaming, video, reviews, forums.

Output: A 3–5 sentence scenario + 3 discussion questions. No answer; that’s the point.

Use when: Standups, calibration sessions, onboarding, or whenever a team needs to align.

8. Single-Case LLM Judge

moderation QA

Purpose: Show how an LLM would call a single piece of content against a single rule, with reasoning.

Input: A piece of content and a specific policy rule.

Output: Verdict (Violating/Not violating/Unclear), confidence, and 3–5 sentences of reasoning.

Use when: Sanity-checking how an LLM judge will behave before deploying it at scale.

9. LinkedIn Prompts

leadership & strategy

Purpose: Pull a random LinkedIn post prompt from a curated deck.

Input: None.

Output: One prompt at a time, from a deck of 100. No LLM call; pure randomizer.

Use when: User wants to post but can’t think what to say.

10. Palate Cleanser

wellbeing & community

Purpose: Reset after a heavy queue.

Input: Toggle: Dogs / Cats / Random.

Output: A random animal photo from a public API. No LLM.

Use when: User has been exposed to hard content and needs a moment.

11. T&S Maturity Check

leadership & strategy

Purpose: Self-assess T&S maturity across six pillars.

Input: 25 multiple-choice questions, scored 1 (Reactive) to 4 (Adaptive).

Output: Overall band, per-pillar scores, and where to focus next.

Use when: User is starting or refreshing a T&S program, prepping a roadmap, or comparing to peers.

12. Confusion Matrix

moderation QA

Purpose: Visualize where labels diverge between actual and predicted.

Input: A CSV with two columns: actual label and predicted (or assigned) label.

Output: A colored confusion matrix plus overall accuracy stats. Runs entirely in the browser.

Use when: QA on a moderation team, evaluating an LLM judge, or comparing classifier outputs.

The prompts themselves

Every prompt that powers a free tool is openly available at /prompts-and-skills. Each prompt is copyable, downloadable as a markdown file, and packaged into SKILL.md bundles for drop-in use with Claude, ChatGPT, or your own agent setup.

Pointers