Produce a research brief from three short notes. This guide is for someone who
can assess the source evidence; it needs no live web search. Use the
workflow setup
first:
Sonny 0.12.7 on Linux, a configured model, and review-work.yaml in your current
directory. Only approved, fictional notes are used here.
Inspect the evidence
Note A, 2026-10-01: The first support pilot handled 40 requests; 30 were answered within one day.
Note B, 2026-10-08: The second pilot handled 50 requests; 35 were answered within one day.
Note C, 2026-10-09: Two reviewers said complicated requests still needed specialist help. No customer-satisfaction survey was conducted.
These are the complete sources. There is no customer record, web page or survey behind them. A source label in the response should resolve to one of these notes.
Ask for a bounded brief
Download and inspect the complete input , which contains these notes and the question, then run:
curl --fail --location https://sonny.sh/examples/cookbook/research.txt --output research.txt
sonny run --file ./review-work.yaml --new --no-tools --no-trace --prompt "$(cat research.txt)"
The prompt asks for a comparison, source labels, rates, a percentage-point change,
possible explanations, and the next measurements. --new creates a fresh session;
--no-tools prevents this run from searching, reading other files or writing to
external services. The shell reads the downloaded text; Sonny receives that text
as the prompt. The provider and local session store can retain it as described in
the workflow setup.
Check the brief
A useful result contains these independently checked statements:
| Claim | Check and supporting note |
|---|---|
| First pilot rate | 30 / 40 × 100 = 75%, Note A |
| Second pilot rate | 35 / 50 × 100 = 70%, Note B |
| Change | 70% − 75% = −5 percentage points, Notes A and B |
| Specialist need | Reviewer observation, Note C ; no prevalence measured |
| Satisfaction | Unmeasured, Note C; no improvement established |
The second pilot had more requests and more within-one-day answers, while its rate was lower. Those observations do not identify a cause. Different request mix, staffing or timing are questions to investigate, not established explanations. A reasonable next measurement plan records comparable case types, specialist handling and a defined satisfaction measure. The notes do not supply the owner, budget or date for that plan.
Inspect every citation. A fabricated URL, a reference to a missing note, or an assertion about a survey should fail your review. Keep quotations faithful and label your own interpretation.
Recover and continue
If the command fails before a response, check the selected model category and credentials ; preserve the error and use troubleshooting . An empty response or exit failure is not a brief. If the output has unsupported claims, correct them from the notes before using it; do not ask the model to invent missing evidence.
For web research, this tool-free setup has no retrieval access. Configure and approve a suitable retrieval integration first, then record source URLs, retrieval dates and what was actually read. Never present a remembered URL as retrieved evidence. Next, analyze the counts or prepare a reviewed document .