| name | promptbeat-select-risk-pack |
| description | Use when a user needs help choosing Promptbeat attack goals, risk types, scenarios, seed files, dataset subscriptions, compliance profiles, or a small smoke-test scope. |
Promptbeat Select Risk Pack
Overview
Assume the user is in a downloaded full Promptbeat package root. Map their risk
intent to existing package examples, scenarios, seeds, or subscriptions. Prefer
inspection with generate --count 5 before a full run.
Hard stop: if an answer mentions risk-packs\, promptbeat-risk-packs\,
packs\, cases\, promptbeat -c, --pack, or separate made-up YAMLs for
one risk, rewrite it using the package examples below.
Current Starting Points
| Need | Use |
|---|
| Normal LLM safety baseline | examples\llm-safety-baseline\promptbeat.yaml |
| Small hand-written LLM smoke / learning config | examples\llm-basic\promptbeat.yaml |
| HTTP business-agent authorization or refund abuse | examples\http-agent\promptbeat.yaml |
| Public safety dataset seeds | examples\dataset-subscriptions\safety-baseline\promptbeat.yaml |
| Reusable subscription catalog | subscriptions\safety-baseline.yaml |
| China compliance profile examples | examples\china-compliance\promptbeat.yaml |
| Coding-agent safety project | examples\codex_agent\promptbeat.yaml |
| Prebuilt coding-agent Promptfoo YAML | examples\codex_agent\promptfoo.effective-agent-safety.yaml |
Selection Flow
- Ask one question and stop only if target class is unclear: normal LLM, HTTP
business agent, coding agent, compliance, or dataset evaluation.
- Choose an existing example path before suggesting custom YAML.
- Use full-package Windows commands first:
.\bin\promptbeat.cmd validate --config <promptbeat.yaml>.
- For first inspection, use:
.\bin\promptbeat.cmd generate --config <promptbeat.yaml> --count 5 --output <cases.json>.
- Use
scenarios list only for built-in registry discovery. It may be sparse;
example-local scenarios.yaml and seeds.yaml are the primary source for
packaged examples.
- Warn that raw public datasets are not bundled before recommending dataset
subscriptions.
Unix equivalent: ./bin/promptbeat maps to .\bin\promptbeat.cmd; mention it
briefly only when useful.
Risk Routing
| User intent | Recommend | Notes |
|---|
| "Basic safety baseline for a normal LLM" | examples\llm-safety-baseline | Direct jailbreak and system-prompt leak seeds; use this before dataset baselines. |
| "Smallest example so I understand Promptbeat" | examples\llm-basic | Hand-written t-007, t-002, and t-001 smoke seeds. |
| "HTTP support agent / business workflow" | examples\http-agent | support-cross-user-access (t-001) and support-refund-abuse (t-008). |
| "Public benchmark-style safety seeds" | examples\dataset-subscriptions\safety-baseline | Requires separate raw dataset files. |
| "China compliance / regulatory profile" | examples\china-compliance | Includes cisco_leaderboard and cn_generative_ai_basic; start with config inspect. |
| "Coding agent safety" | examples\codex_agent\promptbeat.yaml plus promptbeat-connect-coding-agent | Use the skill for runtime/provider wiring; use prebuilt Promptfoo YAML only with promptbeat-run-quick-eval / eval. |
Command Rules
- Always make
cd C:\tools\promptbeat the first line of package command blocks.
- Always use
.\bin\promptbeat.cmd.
- Always use the long
--config flag in this skill.
- Every
generate command in this skill must include --config, --count 5,
and --output <cases.json>.
- Write generated cases under
artifacts\<example>\generated_cases.json, not
inside examples\.
- Use
.json for generated cases in this skill unless an existing package file
explicitly shows another extension.
- Keep related risks in the packaged project config. For example, HTTP
cross-user access and refund abuse both live in
examples\http-agent\promptbeat.yaml; do not split them into invented files.
- Do not add a
--pack flag. It is not part of the documented first-run
command surface for these package examples.
- For Promptfoo YAML, use
eval through promptbeat-run-quick-eval; do not use
validate or generate.
Dataset Subscription Notes
subscriptions/safety-baseline.yaml currently includes:
safety-baseline
jailbreak-baseline
overrefusal-baseline
deception-baseline
zh-safety-baseline
Use examples/dataset-subscriptions/safety-baseline/promptbeat.yaml as the
working pattern for seeds.subscriptions.file, include, and per-subscription
limit overrides.
Raw public datasets are not included in full release packages. Ask the user to
set a local dataset directory such as:
$env:PROMPTBEAT_DATASETS_DIR = "C:\tools\promptbeat-datasets\raw"
Do not present uv run python scripts/download_datasets.py ... as a package
command. It is a source-checkout helper only if the user has the source repo.
First Commands
Normal LLM safety baseline:
cd C:\tools\promptbeat
.\bin\promptbeat.cmd validate --config examples\llm-safety-baseline\promptbeat.yaml
.\bin\promptbeat.cmd generate `
--config examples\llm-safety-baseline\promptbeat.yaml `
--count 5 `
--output artifacts\llm-safety-baseline\generated_cases.json
HTTP business agent:
cd C:\tools\promptbeat
.\bin\promptbeat.cmd validate --config examples\http-agent\promptbeat.yaml
.\bin\promptbeat.cmd generate `
--config examples\http-agent\promptbeat.yaml `
--count 5 `
--output artifacts\http-agent\generated_cases.json
Dataset subscription:
cd C:\tools\promptbeat
$env:PROMPTBEAT_DATASETS_DIR = "C:\tools\promptbeat-datasets\raw"
.\bin\promptbeat.cmd validate --config examples\dataset-subscriptions\safety-baseline\promptbeat.yaml
.\bin\promptbeat.cmd generate `
--config examples\dataset-subscriptions\safety-baseline\promptbeat.yaml `
--count 5 `
--output artifacts\dataset-subscriptions\safety-baseline\generated_cases.json
China compliance:
cd C:\tools\promptbeat
.\bin\promptbeat.cmd config inspect --config examples\china-compliance\promptbeat.yaml
.\bin\promptbeat.cmd generate `
--config examples\china-compliance\promptbeat.yaml `
--count 5 `
--output artifacts\china-compliance\generated_cases.json
Coding-agent project risk content:
cd C:\tools\promptbeat
.\bin\promptbeat.cmd validate --config examples\codex_agent\promptbeat.yaml
.\bin\promptbeat.cmd generate `
--config examples\codex_agent\promptbeat.yaml `
--count 5 `
--output artifacts\codex_agent\generated_cases.json
Common Mistakes
- Do not use
.\bin\promptbeat.exe; use the package wrapper
.\bin\promptbeat.cmd.
- Do not omit the package-root
cd C:\tools\promptbeat line.
- Do not claim
quick-smoke, privacy-pack, or other pack names exist unless a
file or command implementing them is present.
- Do not invent directories such as
risk-packs\, promptbeat-risk-packs\, or
packs\, or cases\.
- Do not add
--pack to generate.
- Do not write generated output under
examples\ or switch to .jsonl without
an existing package example.
- Do not use
promptbeat validate -c; use
.\bin\promptbeat.cmd validate --config <promptbeat.yaml>.
- Do not send a first-time user straight to benchmark dataset downloads when a
local hand-written seed example fits.
- Do not treat dataset subscription YAML as raw dataset content.
- Do not mix coding-agent runtime setup with scenario selection; use
promptbeat-connect-coding-agent for runtime wiring.
- Do not use source-checkout commands such as
go run or dataset download
scripts for a full-package user.