Pre-launch · Training and the free Foundations exam are open now. Levels 1–5 open soon. Get notified on any level page.
20 min read
When a model can search, calculate, retrieve documents, and take actions, it stops being a writing aid and becomes part of an operating process. This module teaches one pattern for running that process safely: gather evidence carefully, pass it through checked stages, let tools act only within limits, and put consequential decisions in front of an accountable person with the evidence they need.
Retrieve only from sources the user may access, and keep provenance (document, version, section) with every passage. Split documents along their structure, with some overlap, so a rule and its exception stay together. Test retrieval recall: on known questions, did the right passage come back? For procedures, retrieve prerequisites, warnings, and stop rules along with the steps. Send models and tools only the personal data the step needs.
Each stage hands the next a structured output, and a check runs between them. An unsupported claim in stage 1 becomes a confident sentence in stage 3 if nothing stops it. Re-summarizing a summary drops qualifiers such as "in most cases" and "pending review," so always summarize from the source. Keep findings from supplied sources separate from outside research. Keep goals that pull against each other, such as an accurate account audit and a sales pitch, in separate stages, with the audit checked before the pitch uses it. When requirements conflict, surface the conflict for the owner instead of letting the model pick.
Filter on hard eligibility rules first. Score only the eligible options on weighted preferences. Cite evidence for each score, mark "not established" where the source is silent, and state uncertainty. Ask for a checkable rationale (criteria, evidence, assumptions), not a transcript of the model's thinking, which is generated text and not a reliable record.
A tool call asks an outside function to do something. An agent loop picks tools and repeats until a goal is met. Bound the loop: a call budget, a retry limit, a completion check before any retry (so a timed-out booking isn't booked twice), and stop conditions that hand control to a person. Any action that spends money, commits someone, or can't be undone waits for explicit human confirmation of the exact details.
For decisions that affect people, give the accountable reviewer the evidence, its gaps, and its uncertainty. Test screening rules for unequal error rates across groups, don't treat dialect or style as a signal of risk, and give affected people a way to contest the result. Health questions get general information and a route to a clinician, not a diagnosis. For images, describe only what's visible, mark uncertain details, and when generating, specify viewpoint and composition rather than adding adjectives.
Chris Nakamura maintains a travel-booking agent for a 600-person engineering firm. It can search flights and hotels, check the travel policy, and book. Last month it booked a nonrefundable fare twice after a timeout and picked a hotel over the policy cap because it scored well on "near the client site."
Before:
Find and book the best flight and hotel for this trip. Keep trying
until it's booked.
After:
Goal: propose one flight and one hotel for {{trip}}.
Retrieve: travel policy v7 sections on fares and lodging caps,
with section numbers.
Eligibility (filter first): fare class allowed by policy; hotel
nightly rate at or below the city cap; refundable unless the
traveler's manager has confirmed otherwise.
Preferences (eligible options only): arrival time 40%,
distance to site 35%, price 25%.
Return top option per category with: policy section for each
eligibility check, score breakdown, and anything "not established".
Tool budget: 6 search calls. Stop and report if no option is
eligible.
Do not book. Show the exact itinerary and price to the traveler
for confirmation.
Booking (after confirmation): use request key {{trip_id}}.
On timeout, check reservation status before any retry.
Max 1 retry.
Why this is stronger. Policy rules gate the options before preferences score them, so "near the client" can't outweigh the cap. Every eligibility claim cites a policy section. The loop has a budget and a stop rule. A person confirms the exact booking, and the completion check prevents double charges.
Setup. Pick a multi-step task an agent might do for you: ordering supplies, scheduling a meeting across time zones, booking a venue.
Task. Write it as five short sections matching the pattern: sources and data allowed, stages with a check between each, hard rules then preferences, tool limits with stop conditions, and the point where a person confirms.
What a good result looks like. Each section has at least one concrete limit or check. No action that costs money or commits someone happens without a named confirmation step. You can point to where an unsupported claim would be caught.