A place to start.
A way to check.

Keep your first task easy to inspect

Complete the quickstart first. Use test data, read the requested destination before approving it, and avoid credentials or external submissions. These prompts intentionally keep the scope small. The model can still make mistakes or try an unexpected action; stop if its behavior differs from your request.

Each task may use your model allowance or incur provider charges. The BotHearth dollar meter estimates work and cannot cap an external bill. For shell or file tasks, enable only the capabilities needed and use a fresh workspace without private files.

Recipe 01 / Browser + file

Read a public page and save a note

Input: the public example page at example.com. Output: out/example.md. You need browser access to that destination and permission to write the result in the task workspace.

Prompt

Open https://example.com/ and read the visible page. Save out/example.md with its page title, source URL, and a one-sentence summary in your own words. Do not follow links, submit forms, or sign in. Finish by naming the saved file.

Check the result

1. Approve the expected https://example.com destination only.
2. Open out/example.md from the task's saved results.
3. Compare the title with the page. The expected title is Example Domain.
4. Confirm the note includes the source URL and describes an example domain used in documentation. Check it contains no invented claims or additional browsing results.

If the file is missing, the task did not satisfy the request. If a different destination appears, inspect why before allowing it. A successful model response without the artifact is not enough.

Recipe 02 / Browser inspection

Review one page you own

Input: a public page you own or have permission to inspect. Replace YOUR_PAGE_URL below with its exact HTTPS URL. Output: out/page-review.md. This exercise inspects one page; it is not a full accessibility or search audit.

Prompt

Open YOUR_PAGE_URL. Record the page title, visible main heading, and the labels and destinations of up to five visible navigation links in out/page-review.md. Identify missing or unclear labels you can directly observe, and explain the evidence for each observation. Do not open other pages, submit forms, or modify the site. Say which checks you could not perform. Finish by naming the saved file.

Check the result

1. Compare the recorded title and heading with the actual page.
2. Manually inspect the listed link destinations and labels.
3. Reject findings that rely on an unseen page, a guessed search ranking, or an unsupported compliance claim.
4. Confirm the report distinguishes observations from recommendations and lists checks it did not perform.

For a site you maintain, use this report as a starting point for a human review. A browser snapshot does not demonstrate screen-reader support, color contrast, or a complete accessibility-tree audit.

Recipe 03 / Local files

Turn a small CSV into a checked summary

Input: three rows of invented reading data. Output: out/reading-summary.md. Use an empty workspace and permit shell/file capabilities for this task. This recipe does not need a website or a login.

Prompt

Create reading.csv in the task workspace using exactly the CSV below. Read it and write out/reading-summary.md with the total minutes and number of entries for each topic, plus the overall total. Do not install packages or contact a website. Keep the original CSV unchanged after creating it. Finish by naming both files.

topic,minutes
history,20
science,35
history,15

Include the CSV immediately after the prompt in the task box. It contains invented data; substitute your real files only after you understand the workspace and provider boundaries.

Check the result

Open both files, check the source rows, and compare the summary with these totals. If the task tries to browse or install anything, stop and inspect that deviation. The input has three data rows; count the header separately.

Help make the next run better

A useful issue includes the BotHearth commit, OS, container runtime, model connection type, a small non-sensitive prompt, expected output, and what actually happened. Include the error text after removing private paths, account data, and credentials.

These recipes are starting points. A reproducible failure is useful evidence; so is a verified result with the exact conditions recorded. Read how to contribute and report security issues privately.