| name | promptbeat-run-quick-eval |
| description | Use when a user has a Promptbeat config or generated promptfoo config and wants to validate, generate a few cases, run a first small evaluation, inspect artifacts, or create a report. |
Promptbeat Run Quick Eval
Overview
Assume the user is in a downloaded full Promptbeat package root. Pick the
correct command path based on input type before giving commands. Project configs
use run; generated promptfoo YAML uses eval.
Command Choice
| Input | Command |
|---|
A. Promptbeat project config such as examples\http-agent\promptbeat.yaml | .\bin\promptbeat.cmd validate --config <promptbeat.yaml> then .\bin\promptbeat.cmd run --config <promptbeat.yaml> --output-dir <dir> |
| B. User only wants generated cases | .\bin\promptbeat.cmd generate --config <promptbeat.yaml> --count <n> --output <cases.json> |
C. Existing generated promptfoo YAML such as generated_redteam.yaml | .\bin\promptbeat.cmd eval --config <promptfoo.yaml> --output-dir <dir> |
| D. Existing result JSON | .\bin\promptbeat.cmd report --input <evaluation_result.json-or-promptfoo-result.json> --output <report.html> |
Safe First Run
- Ask or infer which input type the user has.
- For a Promptbeat project config, validate first, then run with a small output
directory under
artifacts\.
- Use
generate only when the user wants to inspect cases before hitting the
target. State that it writes cases JSON only; it does not evaluate the target
and does not produce promptfoo YAML.
- Use
eval only with an existing generated promptfoo YAML. Add
--provider-file <providers.yaml> if the user needs a provider override.
Direct eval writes Promptfoo artifacts such as promptfoo-result.json; do
not imply it writes Promptbeat evaluation_result.json or report.html.
Do not say promptbeat generate creates this YAML; generate creates cases
JSON only.
- Use
report when the user already has a result JSON: either Promptbeat
evaluation_result.json from run, or raw Promptfoo promptfoo-result.json
from direct eval. --eval-result is also accepted by the CLI, but
--input is the clearest flag for either file.
- Point the user to
evaluation_result.json, report.html, and trace artifacts
only when the command path can produce or has produced them.
Unix equivalent: ./bin/promptbeat maps to .\bin\promptbeat.cmd; mention it
briefly only when useful.
Load Detailed Recipes
Use references/cli-recipes.md for command templates.
Common Mistakes
- Do not use source-checkout commands or
go run in this skill.
- Do not use
.\bin\promptbeat.exe; use the full-package wrapper
.\bin\promptbeat.cmd.
- Do not make
generate a mandatory step before run.
- Do not pass a Promptbeat project YAML to
eval.
- Do not pass generated cases JSON to
eval.
- Do not say
promptbeat generate creates generated promptfoo YAML.
- Do not use
--output with run or eval; both require --output-dir.
- Do not assume a report exists unless
run produced it or report was invoked.
- Do not tell users to rerun
run just because they have direct eval
promptfoo-result.json; report --input can parse it.
- Do not hand-write provider YAML when a placeholder or existing provider-file
example answers the user's question.