Lesson 31 of 32
The Tireless Apprentice
Agents & Research
An agentic or deep-research mode does not answer one question - it takes an errand and runs several steps by itself: planning, searching, reading, comparing, drafting. It works when you supply the three things it cannot supply for itself - a precise goal, limits on what it may trust or spend, and checkpoints where it must stop and show you - and it fails quietly when you do not, returning something confident and well-formatted built on whatever it happened to find first. One rule never bends: the apprentice may draft, fill or prepare, but you send, buy, book and submit.
Two things make a multi-step errand work here. The first is that your standing rules belong in Custom Instructions rather than in the prompt - put "never send, book, buy or submit anything without showing me first and waiting" there once, and it rides along in every chat, including the ones where you forget to say it. The second is to split the errand into two messages: the first asks for a numbered research plan and nothing else, and you correct that plan before any work begins. Switch web search on for anything that has changed recently, and ask for links inline, beside each claim, rather than gathered tidily at the end, so you can see which statement came from which page. For a report you will keep working on, ask for it in Canvas.
From the book Use web search for up-to-date facts and ask it to list its sources with links. Save your standing rules - for example "Never book or send anything without my approval" - in Custom Instructions so they apply to every chat.
- Put "never send or buy without my approval" in Custom Instructions.
- Ask for the research plan first, in a message of its own.
- Request links inline, beside each claim, not pooled at the end.
- Check which connections your app really has before trusting an errand.
- You press Send. Always. No exception for a good draft.
Instead of
Research drip irrigation subsidies in India.
ChatGPT version
**Do this once, in Settings, under Custom Instructions - and every errand you ever send inherits it:** > When a task needs several steps, show me a numbered plan first and wait for my approval. Never send, submit, book or buy anything on my behalf. Put the link beside each fact rather than in a pile at the end. If you could not find something, say so instead of filling the gap. **Then, with web search switched on, send this - and only this:** I am a B.Sc. Agriculture student in Pune. I need the research behind a two-page seminar paper on government subsidies for drip irrigation for small farmers in Maharashtra. Do not research yet. Reply with a numbered research plan and nothing else: which schemes you intend to examine, central and state; what you will establish about each one (who qualifies, what share of the cost is paid, how a farmer applies); and which sites you expect to use for each step. Rules the plan must obey: government portals and recognised agricultural universities first; the publication or last-reviewed date noted for every source, with anything older than two years flagged; no coaching notes, no news articles standing in for the scheme itself, no undated PDF whose issuer you cannot name. I will correct the plan before you start.
Then loop it
- The plan is sound apart from one gap - add a step that sets the small-farmer rules against the large-farmer ones, because that contrast is half my argument. Then run the whole plan.
- The report is far longer than my seminar needs. Give me the five points that matter most, each with its link, and then a separate short list of every place where your sources disagreed with one another.
- Now open the two-page version in Canvas so I can cut it to my word limit myself, and leave a [confirm] marker beside any subsidy percentage that came from a page that was not a government one.
Why it worksGoal, limits and a checkpoint keep the multi-step research on track, and the student approves the plan before the apprentice sets off.
Instead of
Plan a Haridwar trip and book trains.
ChatGPT version
My wife and I are both in our late sixties and live in Ludhiana. Four days in Haridwar and Rishikesh, early February. About ₹40,000 for the whole trip, and neither of us can walk far. **What you must not do:** book, reserve, pay, hold or open a payment page. Not even a free-cancellation hold. You find the options; I decide and I book. Search the web and bring me three separate tables, with a link in every row: 1. Trains from Ludhiana to Haridwar - departure, arrival, class, approximate fare, and whether it arrives at an hour two tired people can manage. 2. Mid-range hotels within a short walk of Har Ki Pauri - price a night, whether the building has a lift, and the real walking distance to the ghat. 3. Temple and Ganga aarti timings for early February, with the page you read them on and the date that page was last updated. Then the four days laid out day by day, with a running total against the ₹40,000 and short walking each day. Write [confirm] against any fare or timing you are not certain of. I would rather check six things myself than discover one invented one at the station.
Then loop it
- The second hotel is the one. Search again for recent guest reviews of it and tell me what people actually say about the lift, the stairs and the walk to the ghat - quote them, and give me the review page.
- Now a one-page printable version in Canvas: the four days, the aarti timing, and blank lines where I will write in the hotel phone number and our coach number once I have booked them myself.
Why it worksThe apprentice searches and compares across many sites, but the clear rule "do not book or pay" keeps the final decision with Harbhajan.
Instead of
Send reminders to everyone who owes me money.
ChatGPT version
**First, a check this errand depends on.** Reading your own mailbox is not the same as searching the web - it needs a connection you have switched on and permitted, and what is on offer differs between apps and keeps changing. Open your own settings and look. If there is no such connection, do not abandon the errand: do the gathering yourself and hand over the result, which is the version below. I run a tiles and sanitaryware shop in Lucknow. Attached is a sheet exported from my own records - customer name, invoice date, amount, and a column marked "payment received: yes / no / not sure". Working only from the rows marked "no" or "not sure": 1. Build a table - customer, invoice date, amount, days outstanding as of today. 2. Then draft one short payment reminder per customer, in simple English, naming that customer's own invoice date and amount. **Send nothing. There is no version of this errand in which you press Send.** Show me the table and the drafts here. I will read each one, correct it, and send it myself from my own account. Where a row says "not sure", write a polite enquiry asking whether payment has already been made, not a reminder. I would rather ask twice than accuse once. And put no UPI id, payment link or bank detail in any draft - I will add those myself if I want them there.
Then loop it
- Two of these were settled in cash and my sheet had not caught up - drop [customer A] and [customer B] from the table and delete their drafts.
- For anything past 60 days, make the wording firmer: the amount, the date, a clear request, no apology. Still someone I want to sell tiles to next year.
- Now write those standing rules as a block I can paste into Custom Instructions - never send mail on my behalf, always show drafts first, never add a payment link - so I do not have to remember them next month.
Why it worksConnected apps let the apprentice gather facts from many emails at once, while the "do not send" checkpoint keeps the human in the loop.
Loop it
With research and multi-step tasks, loop at two points. First, at the plan: read the steps the AI proposes and correct them before it starts - add a missing question, remove an unreliable source, tighten the budget. Second, at the result: check the sources, ask what it could not find and where sources disagreed, then ask for a short summary or table. Never let the loop end with the AI sending, buying or submitting; the last step always belongs to you.