How does the book's canoe-trip prompt look rewritten for 2025 models?

A Beginner's Guide to ChatGPT used a canoe trip to show why context matters: a bare request for camp activities got a generic answer, while a prompt that gave ChatGPT an outdoor expert's role and described the actual day got a useful, tailored one. That lesson has not aged. What has changed is the model on the other side. In April, OpenAI released o3 and o4-mini, reasoning models that think before answering and, for the first time in that series, can use every ChatGPT tool, including web search, data analysis in Python and reading images.

So the 2023 prompt can be rewritten. Not to make it longer for its own sake, but to ask for things an older chat model could not do: look up current information, say where each fact came from, and reason through the trade-offs of a plan.

The method, unchanged

The book's structure still organizes the prompt:

1. Role: who the model should act as.
2. Context: the group, the place, the dates, the gear.
3. Task: what you want produced.
4. Format: how you want it laid out.
5. Iterate: read the answer and refine.
6. Verify: check every safety point against official and local sources.

The 2025 version

Here is a rewrite for a summer trip, written for a reasoning model with search turned on:

Act as an experienced canoe-trip planner for lakes in southern Québec.

Before you plan anything, ask me up to five questions about anything missing that would change the plan. Wait for my answers.

Context: 2 days, 1 night, last weekend of June, lakes with short portages in the Petite-Nation area of the Outaouais. Group: 3 adults and one 12-year-old. Two adults paddle often; one adult and the child are beginners. We own 2 tandem canoes, lifejackets that fit everyone, a tent and a stove.

Task:
1. Search for current official information on required canoe safety equipment in Canada and summarize it as a checklist. Link each item to its source.
2. Draft a two-day plan with distances, rough timings, and a turnaround time for each day.
3. Think through the main risks for this group (summer storms, cold water, a beginner and a child in the boats) and say how the plan handles each one.
4. List what you could NOT confirm online and who I should call to confirm it (park, outfitter, municipality).

Format: a short plan, then a table of risks, then the "could not confirm" list. Mark any statement you are unsure of with [check].

What changed, side by side

That last row comes from o3's image skills: OpenAI says it can work with blurry or low-quality photos, such as a whiteboard or sketch, and zoom or rotate them while reasoning. A photo of the outfitter's map can help it estimate distances. Treat those estimates as rough.

The safety check still belongs to you

A model with search can find good sources and still summarize them wrongly, or find an outdated page. Compare its checklist line by line with Transport Canada's Safe Boating Guide. For human-powered boats such as canoes, the guide lists a lifejacket or PFD (personal flotation device) for each person on board, a buoyant heaving line of at least 15 metres, a bailer or hand pump and a sound-signalling device, with a watertight flashlight and other items required in some conditions. It also explains that if everyone wears a properly fitting lifejacket or PFD, the list of items to carry gets shorter. Check the current rules yourself rather than relying on any summary, including this one.

Two other points from the guide fit a June trip. Summer thunderstorms can build quickly, so if the sky darkens and the wind shifts, head for shore. And cold water shock, the gasping in the first minutes after a fall into cold water, is described as likely more deadly than hypothermia, even for strong swimmers. Then leave a trip plan with someone reliable and call the people who manage the lakes about water levels, campsites and fire rules. A model cannot see this weekend's conditions.

What do reasoning and search add to a planning prompt?

Reasoning adds planning across many constraints at once. A trip plan is a puzzle: daylight, distance, the slowest paddler, a portage, the weather window. A reasoning model works through those before answering, so it is more likely to notice that a beginner and a child should not share a canoe, or that a long crossing late in the day conflicts with a turnaround time. Asking it to "think through the main risks" puts that ability to work on the part of the plan that matters most.

Search adds freshness and traceability. A model writing from memory can only repeat what it learned in training, which may be out of date. With search, it can read current pages and link them, so you can check a claim in seconds instead of hunting for a source. OpenAI says it trained o3 and o4-mini to decide when to use their tools, but naming the sources you want ("official information", "link each item") steers it toward the right kind.

Neither replaces the last step of the book's method. Reasoning can be wrong in a well-argued way, and search can surface an old or local page that doesn't apply to you. The [check] marks and the "could not confirm" list are there to make verification faster. They are a to-do list for you, and the final call on safety stays with official sources and the people who know the water.