executing-plans
executing-plans
Use when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available
Skills 둘러보기
생산성 및 협업 · SKILL.md
targets-scorers
targets-scorersConfigure PyRIT prompt targets, target capabilities, scorers, score aggregation, and target/scorer troubleshooting.
Independent GitHub snapshot · Automated classification · Original documentation governs
Skill 문서
GitHubname: targets-scorers description: "Configure PyRIT prompt targets, target capabilities, scorers, score aggregation, and target/scorer troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT
Targets and Scorers
Use this sub-skill when you need to:
- choose and configure a prompt target
- reason about target capabilities, modalities, auth, and rate limits
- choose or compose scorers
- aggregate, threshold, or batch score results
- debug credential, endpoint, browser, model, JSON, or modality failures
Start here
Fast routing
- Offline dry run, logging, or manual prompt capture:
TextTarget - OpenAI / Azure OpenAI / OpenAI-compatible services:
OpenAIChatTarget,OpenAIResponseTarget,OpenAICompletionTarget - Raw HTTP or API-style endpoints:
HTTPTarget,HTTPXAPITarget - Browser or web UI workflows:
PlaywrightTarget,PlaywrightCopilotTarget,WebSocketCopilotTarget - Provider fan-out or load balancing:
LiteLLMChatTarget,RoundRobinTarget - Local model / torch / download-heavy workflows:
HuggingFaceChatTarget - Exact or regex detectors:
SubStringScorer,RegexScorer - LLM-guided judgment:
SelfAsk*scorers,LlamaGuardScorer,ShieldGemmaScorer - Aggregation, thresholding, and memory-backed evaluation:
TrueFalseCompositeScorer,FloatScaleThresholdScorer,BatchScorer,pyrit.analytics
Boundaries
- For memory and configuration bootstrap, use setup-memory-core.
- For attacks and scenarios that consume targets and scorers, use attacks-scenarios.
- For converters and datasets that shape target input, use converters-datasets.
- For scanner/backend commands, use cli-backend-scanner.
Operating rule
- Pick the smallest target family that fits the endpoint or workflow.
- Check capabilities before assuming the target can accept a modality or conversation shape.
- Prefer local scorers when the signal is lexical or rule-based.
- Use LLM-backed scorers only when the judgment needs model reasoning.
- Keep live-service checks separate from the offline smoke script.