Experience Study Write-Up: A Four-Prompt Claude Chain From A/E Table to Peer Review Checklist

What This Builds

A single Claude conversation that turns your finished actual-versus-expected table into four pieces of paper: a restated table with observations and questions, a memo draft for the reviewing actuary, a plain-language summary for a product or finance reader, and a peer review checklist for the study. Each prompt feeds the next, and you stop at a checkpoint after every one to check the output against your workbook before going on.

The chain handles wording. It never calculates anything official, never explains why a cell came out the way it did, never suggests a change to an assumption and never drafts a conclusion or an opinion. Your workbook is the record. The reviewing actuary decides what the study concludes.

Prerequisites

• A Claude account on a plan your company has approved for this kind of data. The Pro plan starts at $20/month, and the Claude conversation is the only paid tool in the build.
• Total ongoing cost of the finished build: the Claude plan above. Nothing else is needed.
• An aggregated A/E table from your own workbook: a handful of cells with expected, actual and the ratio you computed. No policy-level rows, no names, no identifiers.
• Your company's AI policy read before you paste real results. An aggregated results table is still company data, so this depends on the policy and on an approved tool. The whole chain also works on an invented table with the same shape, which is how to rehearse it first.
• A look at your Claude settings for whether chats are used to improve models and how long they are kept. Consumer plans may differ from business plans, so check the current wording in the product and with your IT team.

The Concept

Think of the chain as a relay race with a referee at every handoff. Each prompt runs one leg: restate, draft, simplify, check. The conversation carries the output forward, so the second prompt can build on the first without you pasting it again. The referee is you. At each checkpoint you compare what Claude produced with the workbook before the baton moves on.

The reason for four small prompts instead of one large one is that errors compound. A figure misread in the restated table would flow into the memo, then into the summary, and by the checklist it would look established. Checking after every step catches it while it is one line.

The prompts also fence the model in. Every prompt tells it to use only what you paste, to write "not stated" when the pasted text is silent, and to stay away from causes, recommendations and conclusions. Claude is good at turning a table into tidy prose. Deciding what a table means is the reviewing actuary's job.

Build It Step by Step

Part 1: Prepare the Table and the Conversation

Open the A/E table in your workbook. Cut it down to the columns Claude needs: a cell label, expected, actual and the A/E ratio you computed. Delete anything that identifies a person or a policy. If your company has not approved pasting real results, replace the figures with an invented table of the same shape.

Open a new Claude chat and choose Sonnet 5 from the model picker. Keep all four prompts in this one chat. The later prompts refer back to earlier output, so a fresh chat would lose the thread.

This guide uses one invented table throughout, for Block A over Period 1. Both labels are invented.

The totals check: 60 + 50 + 48 + 50 is 208, and 57 + 51 + 54 + 49 is 211. The total ratio is 211 divided by 208, which is 1.0144 and is written as about 1.014.

Part 2: Prompt 1, Restate the Table and List Observations

Replace the bracketed text with your own table and send it.

I am documenting an experience study. Below is an aggregated actual-versus-expected table. Use only the figures in it. Do not explain causes. Do not suggest changing any assumption. Do not say what any standard, rule or regulator requires. If something is not in the text I paste, write "not stated".

Tasks:
1. Restate the table as a markdown table with the columns Cell, Expected, Actual and A/E ratio, copying every figure exactly as pasted.
2. List observations, one line per row. For each row say whether the A/E ratio is above or below 1.00 and by how much, working from the pasted ratio only.
3. List up to four questions for the reviewing actuary. Each question must be about the study as pasted. No question may propose a cause or a change.

Table:
[paste your table here]

What good output looks like for the invented table:

• Cell 1: ratio 0.95, below 1.00 by 0.05.
• Cell 2: ratio 1.02, above 1.00 by 0.02.
• Cell 3: ratio 1.125, above 1.00 by 0.125.
• Cell 4: ratio 0.98, below 1.00 by 0.02.
• Total: ratio about 1.014, above 1.00 by about 0.014.

Questions for the reviewing actuary:

1. Cell 3 has the highest ratio in the table, at 1.125. Does the reviewer want Cell 3 handled any differently in the write-up from the other cells?
2. Cell 1 has the lowest ratio, at 0.95. Does the reviewer want Cell 1 handled any differently?
3. The pasted text states no measure of uncertainty for any cell. Does the reviewer want one included, and where would it come from?
4. The pasted text does not state the basis behind the expected figures. Which basis should the write-up name?

Checkpoint 1. Check every figure Claude repeated against your workbook: all four expected figures, all four actual figures, each ratio and the total row. Recompute each "above or below by" figure from the ratio yourself. A wrong digit here would travel through the whole chain. Confirm that no observation or question mentions a cause or a recommendation. If one does, tell Claude which line to remove and ask it to resend the list. Then check it again.

Part 3: Prompt 2, Draft the Memo for the Reviewing Actuary

Fill in the three bracketed statements with your own words. Claude cannot know the purpose, the data or the method, so it must use what you give it.

Using the table and observations above, draft a short memo for the reviewing actuary with these sections: Purpose, Data and period, Method, Observations, Open questions.

Rules:
- For Purpose, Data and period, and Method, use only the statements below. If a section needs something they do not give, write "not stated".
- In Observations, do not write any number. Wherever a figure would go, write a placeholder in the form [figure: what it is], for example [figure: Cell 1 A/E ratio]. Give the direction (above or below 1.00) in words.
- In Open questions, use the questions above without adding new ones.
- Do not explain causes, do not recommend any change to an assumption or method, and do not write a conclusion or an opinion.

My statements:
Purpose: [one sentence on why the study was run]
Data and period: [the block, the period and the source, as you would describe them]
Method: [how you computed actual, expected and the ratio, in your words]

For the invented study, the Observations section of the draft should read like this:

Observations
- Cell 1: the A/E ratio is [figure: Cell 1 A/E ratio], below 1.00. Expected is [figure: Cell 1 expected] and actual is [figure: Cell 1 actual].
- Cell 2: the A/E ratio is [figure: Cell 2 A/E ratio], above 1.00. Expected is [figure: Cell 2 expected] and actual is [figure: Cell 2 actual].
- Cell 3: the A/E ratio is [figure: Cell 3 A/E ratio], above 1.00. Expected is [figure: Cell 3 expected] and actual is [figure: Cell 3 actual].
- Cell 4: the A/E ratio is [figure: Cell 4 A/E ratio], below 1.00. Expected is [figure: Cell 4 expected] and actual is [figure: Cell 4 actual].
- Total: the A/E ratio is [figure: total A/E ratio], above 1.00. Total expected is [figure: total expected] and total actual is [figure: total actual].

The other sections restate your three statements and write "not stated" for anything you left out.

Checkpoint 2. Copy the memo into a document and fill every placeholder from the workbook yourself. For the invented table that means Cell 1 expected 60, actual 57, ratio 0.95, and so on down to the total row. Confirm that every placeholder is gone, that each direction word matches its ratio and that the Purpose, Data and period, and Method sections say what you meant. Search for any sentence that starts to explain a cause or suggest a change, and delete it.

Part 4: Prompt 3, Write a Plain-Language Summary

Write a plain-language summary of this memo for a product or finance reader who is not an actuary. About 120 words.

Rules:
- Use no figures. Describe direction and size in words only. Labels such as Cell 3 and Block A are fine.
- Add nothing that is not in the memo. Do not explain causes. Do not recommend anything. Do not state a conclusion.
- Say plainly that the study is still with the reviewing actuary and that nothing in the summary is a conclusion.
- Avoid actuarial jargon. If a term is needed, explain it in a few words.

A good result for the invented study reads along these lines:

We compared what actually happened in Block A during Period 1 with what we expected, using four groups. Two groups came in slightly below expected and two came in above. The largest gap between actual and expected is in Cell 3. Taken together, actual results were slightly above expected. The study is still with the reviewing actuary, who will decide what it means. Nothing here is a conclusion, and some questions are still open.

Checkpoint 3. Put the summary beside the memo and compare them line by line. Two groups below and two above must match the memo's directions: Cells 1 and 4 are below, Cells 2 and 3 are above, and the total is above. Look for words that sneak in a view, such as "because", "should", "concern" or "good". Check that the summary adds no figure and that the status line about the reviewing actuary is still there. Edit the text yourself rather than asking Claude to guess your intent.

Part 5: Prompt 4, Build the Peer Review Checklist

Build a peer review checklist for this study from the memo's sections: Purpose, Data and period, Method, Observations, Open questions. Add one more section for the plain-language summary.

Rules:
- Two to four checklist items per section, each written as a check the reviewer can tick, in markdown checkbox form.
- Base every item on what the memo says. Do not state what any standard, rule or regulator requires.
- Do not add items about causes, assumptions or conclusions.

The result for the invented study should look like this:

Purpose
- [ ] The purpose statement matches what the reviewing actuary asked for.

Data and period
- [ ] The block, period and source match the workbook.
- [ ] The cell definitions in the workbook match the cells in the memo.

Method
- [ ] Each A/E ratio recomputes from its expected and actual figures.
- [ ] The expected basis named in the workbook is the one the memo describes.

Observations
- [ ] Every figure in the memo ties to the workbook.
- [ ] Each direction word agrees with its ratio against 1.00.
- [ ] The total row recomputes from the cell rows.

Open questions
- [ ] Each question is addressed to the reviewer and none proposes a cause or a change.

Plain-language summary
- [ ] The summary adds no figure and no conclusion beyond the memo.

Checkpoint 4. Add your team's own items. Your team's peer review practice, model governance and the reviewing actuary decide what a review must cover, and the chain does not know any of it. Remove any item that does not apply to your study. Keep the finished checklist next to the memo.

Part 6: Finish and File

Move the finished memo, summary and checklist into your team's usual document locations. The workbook stays the record. Nothing from the chain goes to an underwriter, product manager, client or executive until you have read every line and the reviewing actuary has completed the normal review.

Real Example: The Invented Study Start to Finish

Setup: one Claude chat, four prompts, the invented Block A table above. Input: the five-row table, plus three statements from you for the memo. Output: a restated table with five observations and four questions, a memo draft with placeholders filled by you, a short summary with no numbers and a ten-item checklist.

How the checkpoints work on this example:

• Checkpoint 1 catches a wrong digit. If Claude restates Cell 3 actual as 45 instead of 54, the cell ratio no longer ties: 45 divided by 48 is 0.9375, not 1.125. Your check against the workbook finds it before the memo exists.
• Checkpoint 2 is where the numbers enter the memo, and they enter from your hands. Cell 3 gets expected 48, actual 54 and ratio 1.125 (54 divided by 48).
• Checkpoint 3 holds the summary to the memo's directions. Cells 1 and 4 are below 1.00 and Cells 2 and 3 are above.
• Checkpoint 4 is where the reviewing actuary's expectations come in.

Time saved: the drafting of three documents from one table. The checking is the price, and it is the reason you can trust the result.

What to Do When It Breaks

• Claude explains why a cell is high or low. Reply that the text must contain no causes. Then ask for the section again. If it keeps happening, add the sentence "Do not explain causes" to the end of the prompt and resend. Delete any such sentence before the memo goes anywhere.
• A figure in the restated table is wrong. Stop at Checkpoint 1 and correct it in the chat: say which cell and which figure. Then ask for the table and the observations again. Check the new version against the workbook in full.
• Placeholders are filled with numbers the model chose. Any number in the Observations section of the Prompt 2 draft is suspect. Replace it from the workbook, and tell Claude to use only [figure: ...] placeholders if you rerun the prompt.
• The summary quietly gains a conclusion. Words such as "stable", "concerning" or "on track" are views. Cut them. The summary may describe direction and must never judge it.
• The conversation gets long and later answers drift. Start a new chat, paste the memo you filled in, and run Prompt 4 there. Do not paste the original table again unless you need it.
• Real results went in before the policy was checked. Stop and tell the person who owns your company's AI policy, following your company's process. Use an invented table for rehearsals.

Variations

• Simpler version: run Prompts 1 and 2 only and skip the summary and checklist.
• Extended version: if your company approves it, set the chain up in a Claude Project so the prompts and your team's memo layout are saved. A reviewing actuary's preferred memo headings can go into the project instructions.
• Different study: swap the table for another aggregated comparison of the same shape. The rules in the prompts stay the same.

What to Do Next

• This week: rehearse the chain on the invented table above. Then compare each output with the examples in this guide.
• This month: run it on one real study if your policy allows, and keep notes on where your reviewing actuary edited the draft.
• Advanced: build a short prompt of your own for your team's memo layout and add your team's standard checklist items to Prompt 4.

Advanced guide for actuarial analysts. The chain drafts wording around numbers you computed. It does not calculate the official figures, choose an assumption or word an actuarial opinion. The reviewing actuary decides what the study concludes, and Claude's screens and settings may change, so follow the labels you see.