Lesson 08 · Activity 01

Scope a Discovery Loop

A research loop without limits crawls forever. Here you’ll turn a vague goal into a bounded loop the agent can actually finish — and that you can trust.

Work through it

0 of 5 done

Read the research goal.
Set sources and depth (and what to avoid).
Write a clear stop condition.
List the fields to collect (each with a source link).
Assemble the pieces into one paste-ready instruction, then compare to the model.

The research goal

"Find current grant opportunities that could fund a new community-college STEM outreach program, and put them in one comparison table."

Scope the loop

Fill each box. Saved on this device.

Sources to search (and any to avoid)

Depth per source

Stop condition

Fields to collect for each opportunity

Assemble your launch instruction — combine the boxes above into one paste-ready prompt with its guardrails

Saved automatically on this device.

Reveal a strong model answer

Sources: a handful of named funder databases (e.g., federal grant portals, relevant foundation directories); avoid anything behind a login/paywall or that requires student data. Depth: top ~15 results or first 2 pages per source. Stop condition: "10 qualifying opportunities collected, or all listed sources checked — whichever comes first." Fields: funder, program name, deadline, award amount, eligibility, and a direct source link for every row.

Assembled instruction (copy and adapt):

Search the federal grants portal, two relevant foundation directories, and your state education-agency funding page for grants that could fund a community-college STEM outreach program.
For each result, collect: funder, program name, deadline, award amount, eligibility, and a direct source link.
Stop at 10 qualifying opportunities or when all four sources are checked.
Do not log in or open paywalled pages, and take no other actions.
Flag duplicates and anything past its deadline; note each finding's date.

Takeaway: the stop condition and the per-row source link are what turn an open-ended crawl into a finishable, checkable task — and a paste-ready instruction is the artifact you’ll reuse every time.