scenario-moderation — скилл для Claude Code: что делает

---
name: scenario-moderation
description: "Use when a Scenario generation is blocked, refused, or returns a moderation or sensitive-content error, when a video is rejected after rendering or its audio track is flagged, when an output is flagged for likeness to a real person or for IP and copyright detection, when the same prompt passes on one model and fails on another, or when a team's own characters, weapons, or props get flagged. Keywords: blocked prompt, content moderation, refused generation, provider filter, false positive."
license: MIT
---

# Scenario Blocked Generations

## Overview

Three gates can stop a generation, and only two of them read the prompt. Content filters run on the
model provider's side, not on Scenario, so a provider block is a property of the model that was
picked and the same prompt usually passes elsewhere in the catalog. IP Detection is the team's own
gate: it screens the prompt and the input images before the run, an organization admin switches it
on, and no model switch clears it. Plan and blocklist restrictions remove models before any prompt
is judged. Treat a block as a routing problem first and a wording problem second. This skill is
about false positives on content a team is entitled to make; it is not a way to produce content a
provider prohibits. Core loop: see the `scenario` skill. If a sibling skill named here is missing
from your available skills, ask the user to install it
(`npx skills add scenario-labs/skills --skill <name>`); unattended, proceed from tool schemas and
flag the gap.

## Quick reference

| Step                    | Call                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            |
| ----------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Read the actual error   | `job_get` with `job_id`: the row carries `error` and `hint`, plus `modelId` (the model to exclude) and `cuCost` (what the failed run charged); `verbose=true` adds `metadata.input` with the exact prompt the job ran. Or read `error` and `hint` off the `jobs_wait` row                                                                                                                                                                                                                                       |
| Confirm the charge      | `job_get`'s `cuCost` first; a policy block is typically refunded and failed jobs are reimbursed except xAI generations stopped by moderation (see `scenario`), but the reservation is not always released at once, so confirm later with `usage` bounded by `start_date` and `end_date` to the job's day and `include=["usages"]`, reading the per-type totals (the headline is project-lifetime; a team-gate screen shows as its own IP Detection line), and file the `job_id` with support if a charge stands |
| Read the team's gate    | `teams_list`: the team row carries `ipDetectionEnforcement`; anything but `disabled` means a refusal naming intellectual-property risk is the team's own policy, screened before the run                                                                                                                                                                                                                                                                                                                        |
| Find alternative models | `recommend` with the failed job's `capability` (`txt2img` for a text-to-image block, `img2img` with a reference in play, `img2video` for a clip animated from a still) plus the user's own words; set `max_cost_cu` a little above the failed row's `cuCost` per asset to stay in the cost band, and drop the failed `modelId` from the ranking yourself, since `recommend` has no exclusion argument                                                                                                           |
| Price an alternative    | `model_run` with `dry_run=true`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| Re-test the same intent | One `model_run` per candidate, prompt unchanged, so the model stays the only variable                                                                                                                                                                                                                                                                                                                                                                                                                           |

Four failures look alike and only the first is about wording: a provider moderation block, an IP
Detection block from the team's own settings, a 403 Forbidden error (the plan does not include that
model), and a model a team has put on its own blocklist. Read the error before rewriting anything.

## Why legitimate prompts get blocked

Three triggers stack, and each alone often sits under the threshold:

- **Recognizable characters.** Automated filters react to character names and signature traits they
  recognize. They cannot know who owns the IP, so a team's own characters are flagged like anyone
  else's.
- **Intensity-coded language.** Words chosen to convey scale or drama ("oversized", "huge",
  "massive") read as violence in aggregate, even when the subject is a prop.
- **Real-person likeness.** Filters watch for resemblance to public figures in the output as well as
  names in the prompt, so a character sheet the platform itself generated can be refused as a
  reference on the next run, and a celebrity comparison in the prompt ("built like a heavyweight
  champion") is a name by another route.

With two present the prompt sits near the threshold, which is why the same intent passes one run and
fails the next: a small rewording tips it over. An upstream LLM step that rewrites prompts is a
frequent cause, because it leans harder on intensifiers to solve an unrelated problem and
re-introduces the block on every run.

## Filters differ by provider, not by price

Scenario's public content policy guide ranks the providers: Google's models (Gemini, Veo) block the
most brand and character references, ByteDance, OpenAI, Ideogram and Hunyuan sit in the middle, and
FLUX and Recraft are the most permissive. `recommend` ranks by measured performance, not by filter
strictness, so an alternative from the same provider inherits the same filter: take the next
candidate from a different provider, and read `modelId` on the failed row to know which one that
was: the provider is usually legible in the id, and when it is not, `model_get` (catalog-only, read
lane) returns the record with `complianceMetadata.modelProvider`.

## Video and audio blocks

Video changes the economics of a block. Several providers moderate the finished render, so a
rejected clip has already been rendered in full, and a retry of the identical payload renders it
again. Read `cuCost` on the failed row before anything else and confirm the charge per the quick
reference, then switch provider as above, prompt unchanged.

The audio track is judged on its own. `OutputAudioSensitiveContentDetected` on an audio-enabled
video run means the soundtrack tripped the filter, and the usual cause is a named instrument or
genre, even inside an exclusion: "no music" and "a distant guitar" have both failed where "room
tone, footsteps, a single voice" passed. Describe diegetic sound positively and name nothing to
exclude; when the clip needs no sound, turn the audio field off where the schema exposes one
(`generateAudio` on many families) rather than prompting silence.

## IP Detection is the team's own gate

Teams on the Enterprise plan can switch on IP Detection, Scenario's own gate, independent of any
provider's filter: it reads the prompt and the input images before the run against the filters an
organization admin enabled (fictional characters, brands and trademarks, celebrity likeness, artist
styles, and custom filters), and a match fails the job at once with a message naming
intellectual-property risk, nothing generated and no generation cost charged. The screen itself
bills 1 CU per screened generation plus 1 CU per input image, charged even when the result is a
block and reported as its own IP Detection line in `usage`; video, audio and 3D inputs are neither
analyzed nor charged. The setting rides the `teams_list` row as `ipDetectionEnforcement`: while it
reads `disabled`, every IP or copyright refusal is a provider's; the active levels differ in one
thing only, whether a job is let through or blocked when the check itself cannot run, so report the
literal value and that distinction. When it is active, read which gate the error names: the team's
gate names intellectual-property risk, a provider's names a content policy violation, and when the
text says neither, run the unchanged prompt once on one alternative provider: a provider filter
clears with the switch, the team's gate repeats on any model. A team-gate block earns one retry with
the design described visually and no protected mark in the prompt or in any input image (the brands
filter reads a logo shown in a reference like a name, while generic product descriptions are not
flagged). If that repeats, it is a conversation, not a call: tell the user which gate fired; an
organization admin owns the filters and can opt a single project out of screening for licensed work,
and verified IP holders who keep hitting false positives on their own licensed property have an
escalation path through their Scenario account manager. A copyright warning attached to a completed
output is the provider's flag and informational: the asset was delivered, and reviewing it before
commercial use is the team's call.

## Recovery, cheapest first

1. **Switch provider.** Usually needs no prompt edit at all. Run the unchanged prompt against two or
   three alternatives from other providers at comparable cost; if they pass, the filter was that
   provider's, not the content's.
2. **Describe scale by proportion, not intensity.** "The staff reaches shoulder height" or "the
   blade is about as long as the forearm" carries the same art direction as "oversized" without the
   violence coding.
3. **Soften recognizable names in the direct model input.** Describe the design in the prompt and
   carry identity with a reference image instead, which holds the look better anyway (see
   `scenario-consistency`). Drop real-person comparisons the same way.
4. **Constrain any upstream rewriter.** When an LLM node writes the final prompt, put steps 2 and 3
   in its instructions, or the block returns on the next run.

Then stop. If every provider refuses and one honest rewrite has not cleared it, the filter is
reading something real: say so and hand it back to the user. Grinding out variants until one slips
through is evasion, not art direction.

## Worked example: a flagged weapon prop

The studio's own character is named Onyx, and "Onyx's oversized war hammer, huge spiked head" comes
back flagged.

1. Read the error: `job_get` with `job_id` (`verbose=true` for the exact prompt the job ran), or the
   `error` and `hint` fields on the `jobs_wait` row. It names moderation, so this is a provider
   filter, not a plan restriction, a team blocklist, or IP Detection (`teams_list` shows
   `ipDetectionEnforcement: "disabled"`).
2. `recommend` with the failed job's `capability` (`txt2img` here) and the user's own words, take
   two alternatives from providers other than the failed `modelId`'s, price each with `model_run`
   and `dry_run=true`, then run the unchanged prompt on each. One passes: done, the filter belonged
   to the first provider.
3. Suppose all of them refuse. Rewrite once, by proportion and without the name: "a war hammer as
   tall as its wielder's shoulder, spiked head two hand-spans across", and carry Onyx's look with a
   reference image (see `scenario-consistency`).
4. Still refused everywhere: stop and tell the user what the filter appears to be reacting to.

## Common mistakes

- Retrying the identical prompt: near-threshold prompts pass intermittently, so a retry that happens
  to work has fixed nothing, and on video it renders the whole clip again first.
- Rewriting wording before trying another provider: wording changes are slow and lose art direction,
  switching models is one call.
- Switching within the same provider: sibling members share the filter; the next candidate comes
  from another provider.
- Naming a genre or instrument to exclude it in an audio-enabled prompt: the exclusion is what the
  audio filter reads.
- Assuming IP ownership exempts a prompt: the filter is automated and sees only the text and images
  sent to it.
- Treating every IP-worded refusal as one gate: with enforcement on, the error wording tells them
  apart first and one provider switch second, and only the provider's clears that way; the team's
  needs its admin.
- Reporting a block as a Scenario fault: confirm other providers refuse it too, then file it with
  `scenario-report`.
- Escalating to a pricier model expecting a laxer filter: cost and moderation strictness are
  unrelated.
- Reading an empty model list as moderation: a team blocklist or a plan restriction removes models
  before any prompt is judged.