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.

Reach for it whenWhen the answer needs many steps rather than one - comparing schemes, schools, hotels or suppliers, or building a report with sources - and especially when a mistake buried in the middle of those steps would cost you money, a booking or your signature.

Give each of the chapter's three duties a tag of its own: <goal> for what finished work looks like, <limits> for the sources it may trust and the things it must never do, <checkpoint> for where it has to stop and show you. Put any long material you are supplying - an exported invoice list, a circular, last year's itinerary - at the very top, and the instruction at the very bottom, because the final instruction is the one followed most reliably. Turn web search on yourself; without it you get what the model remembers rather than what is currently true, which on fees and timings is the difference between a plan and a disappointment. Ask for the plan as ordinary chat text and the report as an Artifact, so you argue with the plan and edit the document. When a prompt finally works, its tags belong in a Project's instructions.

From the book Turn on web search so Claude can look up current information and show where it found it. For long research, ask it to put the final report in an Artifact - a document beside the chat that you can edit and download.
Worth knowingClaude is co-operative, and on a long errand that is a hazard rather than a comfort. Tell it to stop at the checkpoint and it stops; tell it to proceed and it proceeds cheerfully past a thin source rather than refusing, so the instruction to mark what it could not verify has to be in the prompt from the start - asked for afterwards, you get a tidy rationalisation instead. Reaching your own mail or files is a separate permission, set up outside the chat, and what is available differs by app and keeps moving, so check your own menus rather than this page. And the chapter's law does not bend here either: it may draft the letter inside the Artifact, but you press Send, you make the payment, you submit the form.
  • <goal>, <limits>, <checkpoint> - and then the instruction, last.
  • Turn web search on; without it you get memory, not facts.
  • Long attachments at the top, your instruction at the bottom.
  • Ask for [unverified] markers in the prompt, not afterwards.
  • Report in an Artifact; plan as chat text you can argue with.
The Open Errand vs. The Bounded Research

Instead of

Research drip irrigation subsidies in India.

Claude version

<who_i_am>
B.Sc. Agriculture student in Pune. This is a seminar paper, so my professor will follow the citations.
</who_i_am>

<goal>
Two pages I can read aloud: the central and state subsidies for drip irrigation available to a small farmer in Maharashtra - which schemes, who qualifies, what share is paid, and the route to applying.
</goal>

<limits>
Sources: government portals and recognised agricultural universities. A coaching note about a scheme is not the scheme, and neither is a news article about it.
Give the publication or last-reviewed date of every source. Flag anything older than two years as possibly superseded.
Where two official pages state different subsidy percentages, show me both and name both pages. Do not average them and do not pick the likelier one.
</limits>

<checkpoint>
Stop before you research anything. Your first reply is a numbered plan only - the questions you will chase, in order, and the sites you expect to use. Then wait for me.
</checkpoint>

<when_you_cannot_find_something>
Write it down as a gap, marked [not found]. A gap I can take to my professor; a plausible invented figure I cannot.
</when_you_cannot_find_something>

<task>
Give me the plan. Once I have approved it, carry it out and put the two-page report in an Artifact, so I can trim it to my word limit myself.
</task>
Open Claude 1,331 characters

Then loop it

  1. The plan is right but for one missing question - add a step comparing what a small farmer is entitled to against what a large farmer gets, because that contrast is the argument of the paper. Now carry it out.
  2. Before I read the whole thing: five lines on what you found, then the places where your sources contradicted each other, then the [not found] list. I will open the Artifact after that.
  3. Revise the Artifact itself rather than posting a new version. Keep the scheme table. Cut the history section. Add the source date beside every percentage.

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.

The Blind Trip vs. The Planned Errand

Instead of

Plan a Haridwar trip and book trains.

Claude version

<travellers>
My wife and I, both in our late sixties, setting out from Ludhiana. We tire on long walks and we do not want a crowded day.
</travellers>

<errand>
Four days in Haridwar and Rishikesh, early February. Find the options. I will choose and I will book.
</errand>

<budget>
About ₹40,000 for everything - travel, rooms, food, local transport. Keep a running total and tell me the moment a choice breaks it.
</budget>

<forbidden>
No booking. No reservation. No payment page. No holding a seat or a room, however free the cancellation. If a step cannot be taken without committing me to something, stop and tell me instead.
</forbidden>

<what_to_find>
Trains from Ludhiana, with the arrival hour judged for two people who sleep badly on a train.
Rooms a short, flat walk from Har Ki Pauri - and whether the building has a lift or only stairs.
Current temple and Ganga aarti timings, with the date on the page you read them from.
</what_to_find>

<honesty_rule>
Mark as [unverified] every fare, timing and distance you could not confirm on a page you can name. I will check those at the counter myself.
</honesty_rule>

<task>
First, three or four lines of your own reasoning about which of the two towns we should sleep in and why - I want to argue with you before you build anything. Then the day-by-day plan with the running total, as an Artifact we can keep editing together.
</task>
Open Claude 1,395 characters

Then loop it

  1. Your reasoning about sleeping in Haridwar persuaded me - keep it. Now look harder at the second hotel: what do recent guests actually say about the lift, the stairs and the length of the walk to the ghat? Quote and link them, and if the only accessibility claim comes from the hotel's own page, say so.
  2. Edit the Artifact into something I can print on one page - the four days, the timings, and blank lines for the hotel's phone number and our coach number, which I will fill in after I have booked.

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.

The Free Hand vs. The Approved Draft

Instead of

Send reminders to everyone who owes me money.

Claude version

[Paste your invoice list here, at the very top - one line per invoice: customer, date, amount, and "paid" or "unpaid" or "not sure". Long material goes first and the instruction comes last, where Claude follows it most reliably.]

<my_shop>
Tiles and sanitaryware, Lucknow. These are customers I want to keep, not debtors I am finished with.
</my_shop>

<how_to_get_the_list>
If I have set up and permitted a connection to my mail, read the last three months of sent invoices and build the list yourself. If I have not - and what is on offer differs by app and keeps changing - work only from the list pasted above. Tell me which of the two you did.
</how_to_get_the_list>

<what_i_want>
A table: customer, invoice date, amount, days outstanding today.
Then one reminder per unpaid customer - four or five lines, simple English, naming that invoice's own date and amount.
For anyone marked "not sure", a polite enquiry rather than a reminder. I would rather ask twice than accuse once.
</what_i_want>

<absolute_limit>
You do not send. You draft; I read every line and send it from my own hand. Include no payment link, no UPI id and no account number in any draft - I will add those myself if I want them there.
</absolute_limit>

<task>
Build the table and the drafts together in one Artifact, so I can correct the wording in place.
</task>
Open Claude 1,342 characters

Then loop it

  1. Two of these paid in cash and my sheet is behind - remove [customer A] and [customer B] from the table and delete their drafts inside the Artifact.
  2. Anything over 60 days: firmer. State the overdue amount plainly, drop the apology, keep the respect. Then tell me if any of them now reads as threatening - I cannot hear my own tone.
  3. Now rewrite my whole prompt, tags and all, with blanks for [month range] and [shop name], so I can keep it in the instructions of a Project called "Shop collections" and run it every 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.