Lesson 10 of 32 · Part One, the principles
Commence with the End in Mind
Describe the finished thing before you ask for it, and put that description last. Lay out the background first - the child, the ground, the figures - then close with one decisive line naming the object you want: a six-week table, a one-page report, a printable checklist. The command read most recently is the one followed most closely, and naming an object rather than a topic gives it something to build towards.
How to use Commence with the End in Mind in Claude
This is Claude's documented behaviour rather than a trick, so the chapter needs no translating: long material first, instruction last. In practice that means tagging the background - <child>, <our_group>, <our_team> - and leaving a single <task> tag at the bottom which describes the artefact. Two features make it stronger. In a Project, the background stops being part of your prompt at all: keep the school timetable, last year's figures or the training constraints as Project files, and each new chat needs only the closing command. And because the artefact is the point of the technique, ask for it as an Artifact - a table you keep editing in place is much closer to the finished object than a table retyped four times in chat.
- Tag the background flat; leave one <task> tag at the very end.
- Describe the artefact in that tag - rows, columns, length.
- Keep repeating background as Project files, not in the prompt.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
This is Claude's documented behaviour rather than a trick, so the chapter needs no translating: long material first, instruction last. In practice that means tagging the background - <child>, <our_group>, <our_team> - and leaving a single <task> tag at the bottom which describes the artefact. Two features make it stronger. In a Project, the background stops being part of your prompt at all: keep the school timetable, last year's figures or the training constraints as Project files, and each new chat needs only the closing command. And because the artefact is the point of the technique, ask for it as an Artifact - a table you keep editing in place is much closer to the finished object than a table retyped four times in chat.
From the book Upload long background files first, then write your final task as the last line of your message. In a Project, keep the background as files so each new chat only needs the closing command.
- Tag the background flat; leave one <task> tag at the very end.
- Describe the artefact in that tag - rows, columns, length.
- Keep repeating background as Project files, not in the prompt.
- Ask for an Artifact so the finished object is editable.
- Say what the artefact must exclude, not only what it holds.
Instead of
Make a study timetable, my son has boards, he is weak in maths.
Claude version
<my_son> [Child's name], class 10, CBSE. Board exams begin in about eight weeks. Strong in social science and English. Weaker in maths and science. He loses marks to careless mistakes more than to gaps in knowledge, so practice matters more than fresh reading. </my_son> <his_week> School from 8 to 2. Tuition on Monday, Wednesday and Friday evenings. He studies best in the mornings and needs a break every hour. </his_week> <what_to_leave_out> No motivational writing. No advice on study technique or sleep. No list of recommended books. </what_to_leave_out> <task> Using everything above, build an eight-week revision timetable as an Artifact: a table with one row per week, columns for Maths, Science, Social Science, English, Practice tests and Rest, extra hours weighted towards maths, and one full rest evening each week. </task>
Then loop it
- Keep the Artifact and add a full mock test every Sunday from week four onwards. Change nothing else.
- Now, using the same background, add a second section to the Artifact: a printable daily checklist for week one that he ticks off himself. One line per task.
- The checklist has drifted into study advice again. Strip it back to tasks only - the <what_to_leave_out> rule applies to every round, not just the first.
Why it worksAll the context comes first and the single clear task comes last, so the timetable is built with every detail fresh in the AI's mind.
Instead of
Write a report for the bank, here are our sales details…
Claude version
<our_group> A women's self-help group of 12 members in a village near Cuttack, Odisha. We make mango and lime pickles. </our_group> <last_year> Sales: [sales figures] Loan: [loan repaid on time] New markets added: [new markets added] </last_year> <our_plan> Buy a sealing machine and sell at the district haat every week. </our_plan> <the_reader> A bank officer deciding whether we can handle a slightly larger loan. He will read this between meetings and wants evidence, not enthusiasm. </the_reader> <what_to_leave_out> No mission statement, no talk of empowerment, no request for the loan amount - he will ask. </what_to_leave_out> <task> Using all of the above, write a one-page report for the bank officer under exactly three headings - Our Work So Far, Our Repayment Record, Our Plan - in clear, simple English. Produce it as an Artifact so I can correct the figures before the meeting. </task>
Then loop it
- In the Artifact, make the plan section concrete: how many more jars a month we expect from the sealing machine, and the reasoning behind that number.
- Keep the English version and add an Odia version below it in the same Artifact, which I will read to the members beforehand.
- It has grown past one page. Cut it back without losing a heading or a figure - the words should go, not the evidence.
Why it worksThe 'Now, write…' line at the end gathers everything above it into one decisive task, so the core request is never lost among the figures.
Instead of
Give a training plan for kabaddi, under-16 boys, ground is small.
Claude version
<my_team> 20 under-16 kabaddi players at a government school in Ludhiana. Weaknesses: stamina in the second half, and poor tackling technique. The boys are returning from holidays, so week one cannot be heavy. </my_team> <our_conditions> A small mud ground. No gym. Five evenings a week, 90 minutes each. The district tournament is in six weeks. </our_conditions> <what_to_leave_out> No supplements and no diet advice - parents decide that. No gym or weights work. No drill that needs equipment we do not have. </what_to_leave_out> <task> Using all of the above, build a six-week training plan as an Artifact: a table with one row per week and columns for Warm-up, Skills, Fitness and Cool-down, each cell giving minutes and the drill, adding to a 90-minute session. </task>
Then loop it
- Week one is still too heavy for boys off holidays. Lighten it in the Artifact and build the load up gradually across the six weeks.
- Keep the table and add three simple tackling drills beneath it, one line each, in plain words I can translate into Punjabi on the ground.
- Now add a single week-one sheet to the Artifact that I can print and pin in the staff room - the table rows for week one and nothing around them.
Why it worksThe coach sets the scene first - players, ground, deadline, limits - and blows the whistle with one clear task at the very end.
Questions
What is Commence with the End in Mind in prompt engineering?
Describe the finished thing before you ask for it, and put that description last. Lay out the background first - the child, the ground, the figures - then close with one decisive line naming the object you want: a six-week table, a one-page report, a printable checklist. The command read most recently is the one followed most closely, and naming an object rather than a topic gives it something to build towards.
When should I use this technique?
When your prompt is long on background and you need one specific finished artefact out of it, rather than an essay about the subject.
How do I apply it in Claude?
This is Claude's documented behaviour rather than a trick, so the chapter needs no translating: long material first, instruction last. In practice that means tagging the background - <child>, <our_group>, <our_team> - and leaving a single <task> tag at the bottom which describes the artefact. Two features make it stronger. In a Project, the background stops being part of your prompt at all: keep the school timetable, last year's figures or the training constraints as Project files, and each new chat needs only the closing command. And because the artefact is the point of the technique, ask for it as an Artifact - a table you keep editing in place is much closer to the finished object than a table retyped four times in chat.
What are the limits of this technique in Claude?
The thing to resist is over-tagging the background into a hierarchy; Claude follows the final instruction reliably, but if the closing task is buried under twelve nested tags it stops being the final instruction in any meaningful sense. Keep the background flat and the task unmistakably last. The other risk is scale: Claude is comfortable with very long background, which tempts people to paste in the entire syllabus or a year of accounts, and the answer then covers all of it. Say in the closing line what the artefact must leave out, not merely what it must contain.
How do I improve the answer after the first try?
When the most important instruction sits at the end, check the first answer against that last line. Did it give the exact thing you asked for - the timetable, the one-page report, the table - or did it wander into the background details? If it drifted, keep your context and simply restate the final task more sharply: "Using everything above, give me only the table." For later rounds, put your change at the end of your message too, so it is the freshest instruction the AI reads.
Loop it
When the most important instruction sits at the end, check the first answer against that last line. Did it give the exact thing you asked for - the timetable, the one-page report, the table - or did it wander into the background details? If it drifted, keep your context and simply restate the final task more sharply: "Using everything above, give me only the table." For later rounds, put your change at the end of your message too, so it is the freshest instruction the AI reads.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.