Lesson 9 of 32 · Part One, the principles
The Ladder of Logical Progression
Step-by-Step Thinking
Split the job into named stages and ask to see the working at each one. Be honest about why: the old advice to add "think step by step" does very little on today's reasoning models, which already reason before answering. What still works is naming the stages yourself - load, then AC load, then comparison - because that is what gives you a rung to point at when a figure is wrong, instead of a single answer you must accept or reject whole.
How to use The Ladder of Logical Progression in Claude
Give each stage its own tag and the loop becomes surgical: you can write "stage two is wrong, redo from there" and the earlier tags are untouched. Skip the incantation - Claude reasons before it answers and telling it to think step by step changes little - but do ask for something it does well, which is to put its assumptions and its reasoning before its conclusion, so you can disagree with the premise rather than the number. Two practical points. Put the figures in a tag of their own and the instruction last, where it is followed most reliably. And ask for the final calculation as an Artifact table, because a reconciliation or a load sheet is a document you will correct several times and download.
- One tag per stage, so you can say "redo from stage two".
- Ask for assumptions and reasoning before the conclusion.
- Ask it to flag any figure I gave that looks implausible.
From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.
Give each stage its own tag and the loop becomes surgical: you can write "stage two is wrong, redo from there" and the earlier tags are untouched. Skip the incantation - Claude reasons before it answers and telling it to think step by step changes little - but do ask for something it does well, which is to put its assumptions and its reasoning before its conclusion, so you can disagree with the premise rather than the number. Two practical points. Put the figures in a tag of their own and the instruction last, where it is followed most reliably. And ask for the final calculation as an Artifact table, because a reconciliation or a load sheet is a document you will correct several times and download.
From the book Ask Claude to show its reasoning step by step, then put the final calculation in an Artifact table you can edit and download. Upload a bank statement or spreadsheet only after removing account numbers.
- One tag per stage, so you can say "redo from stage two".
- Ask for assumptions and reasoning before the conclusion.
- Ask it to flag any figure I gave that looks implausible.
- Put the final calculation in an Artifact table you can download.
- Strip account numbers before uploading any statement.
Instead of
Can this house take one more AC?
Claude version
<my_role> I am a licensed electrician. This is a desk estimate before a site visit, not a certification. </my_role> <the_house> Single-phase connection. Sanctioned load: [kW]. Existing appliances: [list with approximate watts]. The customer wants to add a 1.5-ton split AC. </the_house> <stage_1_present_load> List each appliance with the wattage used. Mark any wattage you assumed rather than took from my list. Total it. </stage_1_present_load> <stage_2_the_ac> State the running load and, separately, the starting load you are assuming for a 1.5-ton split, and say where those typical figures come from. </stage_2_the_ac> <stage_3_comparison> Total against the sanctioned load. Say plainly whether it is comfortable, close to the limit, or over - and when in the day the peak falls. </stage_3_comparison> <before_you_start> If any wattage I have given you looks implausible for that appliance, say so before calculating. I would rather be contradicted than have you reason carefully from my mistake. </before_you_start> <task> Work the three stages in order, showing the arithmetic under each, then list the general points to check on site. Put the load table in an Artifact so I can edit figures as I measure them. </task>
Then loop it
- Your stage 2 figure is high. Redo stage 2 for an inverter AC and carry it through to stage 3 - leave stage 1 exactly as it is. Show what changed and by how much.
- Now update the Artifact table with the inverter figures, and write a simple two-line note for the customer in Tamil on whether to apply for a load increase. No technical words in the Tamil.
Why it worksAsking for each step lets the electrician spot exactly where an assumption was wrong instead of trusting a single yes or no.
Instead of
Why doesn't my bank balance match the books?
Claude version
<context> I am an accountant at a small firm in Pune. Cash book and bank statement for [month] do not agree. </context> <figures> Cash book balance: ₹[amount] Bank statement balance: ₹[amount] Unmatched items I have found: [list of cheques issued but not cleared, deposits not credited, bank charges, interest] </figures> <how_to_lay_it_out> A running table, one row per item: item, nature, add or subtract, amount, running balance, one-line reason. Row one is the cash book balance. One item per row - never net two offsetting items into a single line. Where an item in my list is ambiguous, say so in that row rather than deciding for me. </how_to_lay_it_out> <after_the_table> State whether the difference is fully explained, and if not, the exact amount still outstanding. </after_the_table> <task> Build the reconciliation as an Artifact table I can correct and download. Before the table, two lines on which of my unmatched items you think is most likely to be misclassified, and why. </task>
Then loop it
- You treated the bank charges as a deposit. In the Artifact, fix that row and recalculate the running balance from that row downward only.
- ₹[amount] is still unexplained. List the usual causes of a residual difference in the order I should check them, with what to look at for each.
- Add those checks to the Artifact as a second section, so the whole working and the follow-up live in one document I can download.
Why it worksA running balance after every line makes the error visible at the exact rung where the reasoning slipped.
Instead of
Plan a Kashi trip for us.
Claude version
<us> My sister and I, both in our sixties. Travelling from Madurai to Varanasi and Prayagraj, about seven days, next February. We prefer trains. We cannot walk long distances. We want real time at the ghats and temples, not a sightseeing sprint. </us> <stage_1_routes> The main realistic ways to reach Varanasi from Madurai, each with total door-to-door hours, number of changes, and how much walking or standing it involves. Pros and cons for two women in their sixties. Do not pick one yet. </stage_1_routes> <stage_2_order> For the route you think best, a day-wise order of places across Varanasi and Prayagraj, with your reason for that order rather than the reverse. </stage_2_order> <stage_3_rest_and_booking> Where the rest breaks fall, and what must be booked early and how far ahead. </stage_3_rest_and_booking> <when_unsure> Mark any timing, fare or temple arrangement you are not confident about as [check]. A gap in my notes is fine; a wrong darshan timing is not. </when_unsure> <task> Write your reasoning for each stage before that stage's conclusion, so I can reject stage one without losing the rest. Then produce the plan as an Artifact. </task>
Then loop it
- Your stage 1 reasoning underrates how tiring an overnight connection is at our age. Redo stage 1 with a one-night halt option added, and compare the two on the same criteria. Hold stages 2 and 3 until I choose.
- We will take the halt. Now rewrite the Artifact as a simple day-by-day table, and add a short list of questions to ask the travel agent.
- Good. Rewrite my whole prompt, stage tags and all, with blanks for [destinations] and [month], so I can keep it in a Project for our next yatra.
Why it worksBreaking the trip into route, order and rest lets her question each decision before it is buried in a finished itinerary.
Questions
What is The Ladder of Logical Progression in prompt engineering?
Split the job into named stages and ask to see the working at each one. Be honest about why: the old advice to add "think step by step" does very little on today's reasoning models, which already reason before answering. What still works is naming the stages yourself - load, then AC load, then comparison - because that is what gives you a rung to point at when a figure is wrong, instead of a single answer you must accept or reject whole.
When should I use this technique?
When the answer rests on a calculation or a chain of assumptions - an electrical load, a reconciliation, a route - and you need to audit the reasoning rather than trust the conclusion.
How do I apply it in Claude?
Give each stage its own tag and the loop becomes surgical: you can write "stage two is wrong, redo from there" and the earlier tags are untouched. Skip the incantation - Claude reasons before it answers and telling it to think step by step changes little - but do ask for something it does well, which is to put its assumptions and its reasoning before its conclusion, so you can disagree with the premise rather than the number. Two practical points. Put the figures in a tag of their own and the instruction last, where it is followed most reliably. And ask for the final calculation as an Artifact table, because a reconciliation or a load sheet is a document you will correct several times and download.
What are the limits of this technique in Claude?
Claude will not fight you on a figure you supplied. Give it a wrong wattage and it will reason impeccably from the wrong wattage, and the confident presentation makes the error harder to spot, not easier - so ask it explicitly to flag any input that looks implausible before it starts. Watch the arithmetic too: visible working is auditable, not verified. And before uploading a bank statement or a customer's electrical schedule, strip account numbers and personal details; the staged format tempts people to paste the whole document in when a list of figures would do.
How do I improve the answer after the first try?
Step-by-step answers are easy to loop because you can point at a single rung. Read each step and ask: Is this number right? Is this assumption true for me? When one step is wrong, do not ask for a fresh answer - say "Step two is wrong because…; redo from step two." The steps before it stay, and the fix flows forward. For money, health or safety work, treat the steps as a draft to check yourself or with a professional, never as the final word.
Loop it
Step-by-step answers are easy to loop because you can point at a single rung. Read each step and ask: Is this number right? Is this assumption true for me? When one step is wrong, do not ask for a fresh answer - say "Step two is wrong because…; redo from step two." The steps before it stay, and the fix flows forward. For money, health or safety work, treat the steps as a draft to check yourself or with a professional, never as the final word.
This lesson on the other tools
The technique is the same; the way you write it is not. ChatGPT, Gemini.Perplexity.